Why the benchmark works this way

Principle 8 of the spec says record all assumptions so the model can later be challenged and revised. This file is that record.

Each decision states what we chose, why, what it costs us, and what evidence would overturn it. A decision with no overturning condition is a belief, not a decision, and does not belong here.

Append new decisions. Do not edit a decision in place: supersede it with a new one and mark the old as superseded, so the reasoning stays auditable.

D1: Nine dimensions, no headline ranking

Choice. Report nine 0–100 dimension scores. Never compute a composite.

Why. A composite collapses exactly the information the benchmark exists to produce. Two countries with the same average can have opposite shapes, and the shape is the finding.

Cost. Harder to communicate. There is no number to put in a headline.

Overturned by. Evidence that the nine dimensions are so correlated that the shape carries no information beyond a single factor. Watch diagnostics.dimensionPairs: if most pairs sit above r ≈ 0.9 on a larger country set, the dimensional structure is not earning its keep.

D2: Normalisation is relative to the country set, not to an absolute frontier

Superseded by D16 on 2026-08-26. The relative principle stands. What changed is which set defines the scale: a fixed reference set rather than whichever countries happen to be loaded.

Choice. Min-max each indicator across the countries in the run, mapping the weakest to 0 and the strongest to 100.

Why. No defensible absolute frontier exists for most of these indicators, and inventing one imports a value judgement we cannot support.

Cost. This is the single most consequential assumption in the model. Scores from two different country sets are not comparable. Adding an eleventh country changes every number. A country scoring 0 is the weakest of the set, not incapable.

Overturned by. A move to a large enough country set that absolute anchoring becomes possible, or an explicit decision to anchor against fixed reference values per indicator.

D3: Equal weights inside a dimension

Choice. Every indicator in a dimension counts the same.

Why. Any other weighting in v0 would be arbitrary, and arbitrary weights are harder to challenge than equal ones because they look considered.

Cost. A dimension carried by four indicators gives each 25% of the score regardless of quality. A weak indicator drags as hard as a strong one.

Overturned by. Delphi construct-validity ratings that are stable across several real panels. Weight by panel-rated validity only once the panel itself has been shown to agree.

D4: Confidence is reported beside the score, never inside it

Choice. confidence = coverage × recency × source_quality, published as its own number. It never adjusts, discounts or shrinks the capability score.

Why. Folding evidence quality into a capability score conflates two different claims: how capable a country is, and how well we know. A reader can weigh those separately; a blended number hides the trade.

Cost. Two numbers to carry everywhere. Consumers who want one number will invent their own blend, probably worse than ours.

Overturned by. Nothing we can foresee. This is close to load-bearing.

D5: Missing data is dropped, never imputed

Choice. A missing indicator leaves the dimension mean and lowers coverage.

Why. Principle 6. Imputation makes a thin evidence base look identical to a thick one.

Cost. A dimension with one observed indicator out of eight still produces a score, computed from that single indicator. The score looks as solid as any other; only the confidence number reveals it is not.

Overturned by. Nothing. But see D6 — this is why a floor may be needed.

D6: Dimensions are scored at any coverage above zero

Choice. As long as one indicator is observed, a dimension gets a score.

Why. Suppressing low-coverage dimensions would blank out Experimentation entirely and hide the finding that it cannot be measured.

Cost. This is a live risk. Experimentation is scored from two indicators out of eight and reads as a real measurement in the flat table. The v0 in-session estimates disagree with it by up to 56 points.

Overturned by. A published deliverable. Before anything is published, either introduce a coverage floor below which a dimension reports null, or mark low-coverage cells visually in every output. Open, unresolved.

D7: Winsorize with Tukey fences at k = 3

Choice. Clip values beyond the quartiles ± 3 IQR. Record which values were clipped.

Why. The spec says winsorize if necessary. With ten countries a percentile rule such as p5/p95 would always clip the top and bottom country, destroying the variation the benchmark exists to expose. Tukey at k = 3 clips genuine extremes and usually nothing at all.

Cost. On heavy-tailed indicators such as patents, one country can still drag the whole scale.

Overturned by. Evidence that a specific indicator's distribution needs a transform rather than a clip. Prefer adding a transform to the registry over lowering k.

D8: Only the most recent observation, no trends

Superseded by D22 and D24 on 2026-08-26. The scoring rule stands: a score still uses only the latest observation and nothing is imputed. The "no trends" half fell when momentum shipped as a separate layer, on a matched basket against the current frame.

Choice. One value per indicator per country: the latest non-null year.

Why. v0 is a structural test, not a time series. Trends multiply the methodological surface before the cross-section is defensible.

Cost. A country improving fast and a country decaying fast look identical. Capability arguably lives in the derivative.

Overturned by. The cross-section holding up. Trend is the obvious v1 extension and the data is already fetched from 2000 onward.

D9: Gap indicators stay in the registry

Choice. Indicators the spec asks for and no dataset supports are recorded with ingest: 'gap'. They lower coverage, appear in the data-gap report, and are shown to the Delphi panel.

Why. They are the collection agenda, and deleting them would make the confidence scores lie.

Cost. Confidence numbers look bad. That is the point.

Overturned by. Nothing. Do not delete gaps to make numbers look better.

D10: Inspectability is a hard filter on sources

Choice. Reject sources whose underlying data or method cannot be inspected, even when they cover every country. This is why university-industry collaboration and volunteering are gaps despite GII and CAF publishing figures.

Why. Principle 4. A benchmark built on unauditable inputs cannot be challenged, and the whole design assumes challenge.

Cost. Real coverage loss in Coordination and Shared Purpose.

Overturned by. A source opening its microdata, or an explicit decision to accept composite indices with a recorded quality penalty via source.tier.

D11: Delphi output never enters the capability score

Choice. Panel estimates live in delphiScore / delphiIqr. score is indicator-derived only. blendedScore falls back to the panel only where no indicator evidence exists at all, and blendedFrom records which was used.

Why. Mixing model judgement into an indicator score makes the result unauditable and makes the panel's disagreement with the indicators invisible — and that disagreement is the most useful output of the panel.

Cost. Consumers must decide for themselves what to do with two numbers.

Overturned by. Nothing at v0. Any future blending must be a new, explicit, named field, never a change to score.

D12: Panel disagreement is recorded, not averaged away

Choice. Keep the median and the interquartile range. Flag IQR > 25 as dissent. Panelists are instructed not to converge for the sake of converging.

Why. A stable disagreement is a finding about the dimension. Classical Delphi pushes toward consensus; we want the residual.

Cost. No single clean number per cell.

Overturned by. Evidence that panel dissent is noise rather than signal — for instance if dissent does not correlate with low coverage across several runs.

D13: Panel diversity comes from vendors, not from model size

Choice. One top-tier model per vendor, paired with a fixed analytical stance. Do not add cheaper or smaller models to widen the panel.

Why. Delphi needs independent error. Two models from one lab share training data and agree for reasons unrelated to the evidence. A weaker model adds variance that reads as disagreement but is just incapacity, and the IQR is the output we care about most.

Cost. None worth counting. pnpm bench cost puts a four-panelist, two-round, ten-country run in single-digit dollars. Cost is not a constraint at this scale and must not be treated as one.

Overturned by. Evidence that stance dominates model, in which case several stances on one model would be as good and simpler to reason about.

D14: Provenance is stored, never inferred

Choice. Every Delphi run file declares provenance: gateway, in_session, human or mock. Downstream code branches on that field.

Why. Provenance was previously inferred by string-matching the model name. That is how a dry run ends up quoted in a report.

Cost. Hand-authored runs must remember to set it. pnpm bench validate catches it when they do not.

Overturned by. Nothing.

D15: The World Bank is the only wired ingestion source in v0

Choice. One adapter. Everything else is a gap.

Why. One well-understood source with a documented API beats four half-understood ones, and it makes the gap list honest rather than a mixture of "no data exists" and "we did not get to it".

Cost. Several gaps are gaps only because no adapter exists, not for any methodological reason. Those are marked in their registry notes. The strongest candidates are OpenAlex citation impact, V-Dem civil society and polarisation, and UNCTAD export concentration.

Overturned by. Writing the next adapter. Each one is independent work.

D16: The normalization frame is pinned to the ten reference countries

Supersedes D2. Recorded 2026-08-26, when the first six countries were added beyond the prototype set.

Choice. Every indicator's Tukey fences and its 0 and 100 endpoints are computed from the ten reference countries only. Every other country is scored against that same fixed frame. A country outside the frame clamps to 0 or 100 and the cell is flagged outOfFrame.

Countries carry frame: 'reference' | 'extended' in packages/core/src/model/countries.ts. The reference ten are the original prototype set.

Why. Under D2 the scale was recomputed from whichever countries were loaded, so adding one country silently moved every existing score. That makes the benchmark unusable as an ongoing instrument: no published number survives the next data load, and no two runs are comparable. Since countries and indicators will keep being added, the scale has to hold still.

Verified when the six Latin American countries were added: 0 of 90 reference cells moved.

Cost.

• The reference set is now load-bearing and effectively frozen. Changing its membership rebases everything, and that is a deliberate, announced act.
• An extended country genuinely outside the reference range loses information at the clamp. Colombia already sits near the floor on Trust. If clamping becomes common the frame is too narrow for the countries being asked about, and that is the signal to rebase rather than to widen quietly.
• The ten reference countries are not a representative sample of the world. They were chosen to expose different capability structures, so the frame is biased toward the contrasts they happen to span.

Overturned by. A sustained pattern of outOfFrame cells, or a decision to move to absolute anchoring per indicator. Either way, rebasing is a versioned event: bump a frame version, re-publish, and say plainly that the old numbers are not comparable.

D17: Confidence bands are fixed thresholds, and not a red-to-green scale

Recorded 2026-08-26.

Choice. Four bands in packages/core/src/pipeline/confidence.ts: very thin below 0.25, thin to 0.45, usable to 0.65, good above. The viewer colors by band on an ordinal ramp from muted navy to brand lime. The report prints the band label. One source of truth, so the two cannot drift.

Why. Confidence is a product of three fractions, so its usable range is compressed and small differences near the bottom matter more than they look. Bands make that legible where a raw number does not.

Red to green was rejected: it fails for the most common colour vision deficiencies, and it reads as pass and fail when the thing being encoded is a quantity. The chosen ramp was checked with the palette validator. The worst adjacent pair separates at dE 18.7 in light and 18.2 in dark under normal vision. The dark pair sits at dE 6.8 under tritanopia, which is acceptable only because the numeric value is printed beside every bar and the bar length carries the same magnitude.

Cost. The thresholds are a judgement. Nothing in the data says 0.45 is the line between thin and usable.

Overturned by. Evidence about how readers actually act on the bands, or a change to the confidence formula that shifts its range.

D18: One display for every 0 to 100 score

Recorded 2026-08-26.

Choice. Every 0 to 100 number in the viewer renders through a single component, Score, as a filled chip carrying the number and coloured by one of four bands from packages/core/src/pipeline/bands.ts. Dimension scores, panel medians and normalized indicator values all use it. No table gets its own treatment.

Why. The first version tinted the cell background by value at low opacity. Across a sixteen-country table you could not tell 1.1 from 98.7 at a glance, which is the only reading that table exists to support. Three different renderings of the same quantity had also accumulated: a tint, a plain number, and a bar.

Four discrete bands rather than a continuous ramp, for the same reason Metacritic uses bands: a continuous tint cannot separate 20 from 40 at chip size. The ramp recedes into the page at the bottom and reaches brand lime at the top, in both themes, so low values sink and high values stand out. Not red to green, for the reasons in D17.

Worst adjacent pair separates at dE 22.6 light and 22.0 dark under normal vision, and every label clears 4.5:1 against its fill.

Cost.

• The band edges at 25, 50 and 75 are arbitrary. Two countries either side of a boundary look further apart than they are, which is the standard cost of banding and the reason the number is always printed inside the chip.
• The labels say "strong" and "weak", which sound absolute. They are positions against the reference frame, and the tooltip and the legend say so.

Overturned by. Evidence that readers misread band edges as real differences, or a move away from a frame-relative scale.

D19: Extended countries get no visual marking

Recorded 2026-08-26. Reverses a choice made the same day.

Choice. Countries added after the reference frame was fixed are displayed exactly like the reference ten. No badge, no marker, no dimming.

Why. A marker was briefly shipped. It implied the numbers were less trustworthy, which is false: every country is measured against the same frame by the same method, and that is the entire point of D16. The distinction is real but it is about how the scale was built, not about the quality of any country's score, so it belongs in the method documentation rather than on every row.

Cost. A reader cannot tell from the table which countries defined the frame. The method page and D16 say which ten they are.

Broken and restored, 2026-08-26. The comparison panels added by D30 printed the word frame beside every reference country in their lists, which is the same marking this decision removed, inverted. It is gone again. The distinction stays where it belongs, in the method page and in the glossary.

Overturned by. A case where the distinction changes how a number should be read, most likely a country clamping at 0 or 100. Flag outOfFrame on that cell rather than reinstating a badge on the country.

D20: Documented deliveries are recorded as evidence and never scored

Recorded 2026-08-26.

Choice. data/evidence/records.json holds evidence records: a documented case of a country doing the thing an indicator is meant to measure, filed against an indicator whose ingest is gap. Each record carries a claim, one published number with its reference period, a source with a tier and a retrieval date, and a required limits field saying what the case does not show.

Records never enter DimensionResult.score and never raise confidence. They are schema-checked by pnpm bench validate and displayed on the country page under the dimension they belong to.

A gap becomes an indicator only when a comparable series covers at least two reference countries, which is the minimum buildFrame needs to produce a scale. Promoting one is a separate, recorded act.

Why. Brazil scores 11.5 on Building. Every measured indicator in that dimension is industrial output: manufacturing value added, high-technology exports, labour productivity, economic complexity. The dimension is defined as the capacity to build and deliver, and it currently cannot see a national programme that was delivered. Pix and GOV.BR are exactly that, and both sit inside large_project_delivery, which is a declared gap.

The alternative was to put those cases in the page copy above the chart. That was rejected. Curated national successes with no schema, no source discipline and no limits can be assembled for any country, and a page that argues against its own number is unfalsifiable. Encoding the cases makes them checkable, keeps the score honest and turns the disagreement into a measurement task.

The two constructs inside Building are now named: industrial output, which the six measured indicators carry, and state delivery capacity, which large_project_delivery carries and no dataset covers. The nine dimensions do not change. See KNOWN-ARTEFACTS A11.

Cost.

• A record is hand-written, so the layer is only as good as its author. The required limits field and the validator reduce that risk without removing it.
• One country with two records and fifteen countries with none reads as a claim about Brazil. It is a claim about what the indicators cannot see, and the same cases exist elsewhere. Estonia, India and Uruguay have obvious candidates and no records yet.
• The layer can become a place to park advocacy. The promotion rule is what stops that: a record is a step toward an indicator, and it is not a substitute for one.

Overturned by. A comparable delivery series covering the reference set, at which point large_project_delivery leaves gap, the records become source notes on a scored indicator, and this decision is superseded.

Update, same day. The layer opened with two Brazilian records, which read as a claim about Brazil. It now holds 15 records across four countries.

Brazil has 11, covering half a century: Embrapa from 1973, Proalcool from 1975, the immunisation programme from 1973, SUS from 1988, electronic voting from 1996, Plano Real in 1994, deepwater and pre-salt oil, Bolsa Familia and its registry from 2003, Luz para Todos from 2003, GOV.BR from 2019 and Pix from 2020. Estonia has X-Road, India has the Jan Dhan accounts, and Uruguay has the renewable electricity build and Plan Ceibal.

Records also stopped being a single-indicator affair. Twelve sit on large_project_delivery, two on institutional_responsiveness (Plano Real, Proalcool) and one on government_foresight_capacity (Embrapa), so the layer documents three gaps in three dimensions.

Brazilian sources arrived with the second batch. Seven records now cite IPEADATA, which serves series from the Ministry of Social Development, the Superior Electoral Court and the National Petroleum Agency at official_statistical tier, against the international_organization tier of the World Bank records. That is the per-point tier from D25 earning its place.

Several records exist to keep the layer from becoming advocacy. The immunisation record carries a capability that eroded, from 99 percent coverage in 2003 to 68 percent in 2021. Plan Ceibal carries a delivery whose only available measure is weak and stale. The pre-salt record states that individual platforms ran years late, which is precisely the cost and schedule evidence the indicator wants and the production figure hides. A library of national wins would include none of those sentences.

Fifteen records are still not a series, so the promotion rule above is unchanged.

D21: GEM is wired by hand, venture capital stays a gap

Recorded 2026-08-26.

Choice. Two Experimentation indicators leave gap and become ingest: 'manual', entered in data/observations/manual.json from the GEM Adult Population Survey key indicators:

• early_stage_entrepreneurial_activity, the TEA rate as published.
• failure_tolerance, stored as 100 minus the published fear of failure rate.

Sixteen countries, 15 of them from 2023 or later. Singapore last took part in 2014 and is entered at that year, so the recency term marks it down to a weight of 0.17 rather than hiding it.

venture_capital_gdp stays a gap. The OECD SME and Entrepreneurship Financing scoreboard is the only inspectable aggregate. Checked on 2026-08-26, it carries venture capital for 6 of these 16 countries, in national currency rather than as a share of GDP, with 2022 as the latest year. Brazil, India, South Africa and Singapore are absent. Wiring it would score the rich half of the set and lower coverage for the rest, which is the wealth-proxy failure of A3 rebuilt on purpose.

Why. Experimentation was the worst-measured dimension in the benchmark and A1 said not to publish it. Two of eight indicators carried it, both counting formalised invention, which is close to the opposite of the many-cheap-attempts construct.

What changed:

• Experimentation confidence rises from 0.178 to 0.394 for most countries, which moves it from very thin to thin.
• Correlation with log GDP per capita falls from 0.624 to 0.568. TEA itself correlates at −0.296, the first direct capability measure in the registry that does not track income upward.
• Korea's 100 disappears. It scores 75, because the patent filing artefact is now diluted by two indicators where Korea is ordinary.
• Mean distance between the panel and the indicators falls from 27.1 to 18.3 points across the sixteen countries.
• No other dimension moved. Verified against the previous scores.json.

Cost.

• Hand-entered data goes stale silently. There is no fetcher, so somebody has to re-export from GEM and update the file. The retrieval date is on every observation.
• TEA counts every new business, so necessity self-employment weighs the same as a funded startup. Chile at 29.4 and Argentina at 21.4 sit above the United States at 17.7 for that reason. This is a real property of the measure and it is now in the indicator notes.
• GEM publishes the fear of failure rate among adults who already see good opportunities, so the denominator is not the adult population the unit label suggests.
• Four of eight Experimentation indicators are still gaps, so the dimension is better measured and not well measured.

Overturned by. An inspectable venture capital or business R&D series that covers the reference set, or a GEM licensing change that stops the data being usable this way.

D22: Momentum is measured on one ruler and a matched basket

Recorded 2026-08-26.

Choice. Ingestion keeps every year the World Bank returns from the start year rather than the latest value alone. Scoring still reads only the latest value, so no score moves. On top of that history, every country and dimension carries a momentum object: the change in score over ten years, plus the yearly series behind it.

Two rules make the number mean something.

1. One ruler. Historical values are normalised against the frame built from the reference countries' current values. The scale does not move, so a change in the score is a change in the country. A value outside today's frame clamps and the clamp is counted in momentum.clamped. 2. A matched basket. Only indicators observed at both ends of the span enter, and that same basket is used for every year in between. A dimension that gained an indicator would otherwise show movement belonging to the dataset. The basket is reported with its size, and baseScore and currentScore describe the basket rather than the headline score.

A trend is reported when the basket holds at least two indicators and covers at least half of what the country is currently scored on in that dimension. An observation older than five years does not count toward a year.

Why. Capability is a rate as much as a level. Every existing index publishes levels, and levels on public data reproduce the development ranking, which is the failure this benchmark exists to avoid. Direction is where the analytical value is, and the data to compute it was already being fetched and thrown away.

The first run says something the levels cannot. Brazil sits near the floor on Agency at 31 and has moved +26.2 points in ten years against a median of +12.7. Coordination fell in 11 of 16 countries. Estonia lost 14.2 points on Adaptability. India lost 19.3 on Learning.

Cost.

• The observation file grows from 203 KB to 3.6 MB. It is still plain JSON and still inspectable.
• Momentum and score sit on different baskets, so the two numbers are not directly comparable. Every surface that prints a trend prints the basket size next to it.
• Several indicators track worldwide technology diffusion, so almost every country rises on Anticipation and Agency. A positive number is not evidence of catching up. The report prints the median change per dimension for that reason.
• Ten years is a choice. A longer span covers fewer indicators and a shorter one is mostly noise.
• Doing Business indicators are frozen at 2019, so any basket containing them measures a shorter period than it claims.

Overturned by. Enough indicator history to score a full basket at both ends, which would let momentum use the whole dimension rather than a subset. A move to absolute anchoring would also change what a fixed ruler means and this decision would need restating.

D23: The perception layer is retired, and the cost is visible

Recorded 2026-08-26.

Choice. Seven indicators leave the scored set. They keep their rows, with ingest: 'retired', which means a dataset exists and this project rejected it. Retired indicators are not fetched, not scored, and they lower coverage exactly as a gap does.

One indicator is added: intentional homicide rate, from the World Bank, in Trust. It is counted by police and health systems rather than reported as an opinion.

Why. Grouping every scored indicator by its own measurement class showed that the wealth correlation was a property of the evidence, not of the dimensions. Perception indicators averaged 0.75 against log GDP per capita with 75% above the wealth-proxy line. Direct capability measures averaged 0.55 with 20% above it. The benchmark exists to avoid reproducing a development ranking, and the perception layer was the mechanism reproducing it.

Effect on the wealth correlation:

Coordination drops below the wealth-proxy line. The anticipation and coordination pair, which correlated at 0.94 and was the strongest duplicate candidate in the model, is no longer flagged. One duplicate pair remains, anticipation and agency at 0.94. One perception indicator remains in the whole scored set: GEM fear of failure, which is a self-report about the respondent rather than a judgement about the country.

Cost. This is the important part.

• Coordination now rests on one indicator, time to export, frozen at 2019. Confidence falls from 0.46 to 0.079. Trust rests on two and falls to 0.20. Shared Purpose rests on two and falls to 0.25. Those three dimensions are effectively unmeasured, and the scores they print should be read as such. The radar draws them dashed, the tables print the band, and the numbers still move enough to mislead a casual reader. That is a real risk this decision accepts on purpose, because the alternative is a number that looks solid and measures income.
• Estonia scores 97.8 on Coordination and South Africa scores 0.0, both on a single 2019 border-time measure. Neither is a finding.
• The replacement is not income-free. Homicide correlates with log GDP per capita at 0.77 in this set, above the threshold. The difference is that the mechanism is visible and arguable rather than definitional.
• Momentum baskets shrink where retired series carried them.

Overturned by. Observable replacements: court throughput and case clearance, budget execution rates, cross-agency programme delivery, voter turnout, volunteering rates, civic participation. Each one that lands raises the coverage this decision knocked down. If none land, the honest conclusion is that Coordination and Trust cannot be measured with public data, and they should be reported as unmeasured rather than scored.

D24: Two spans for a dimension, and a full line for every indicator

Recorded 2026-08-26. Extends D22.

Choice. Three changes to the trend layer.

1. Ingestion defaults to 1990 rather than 2000. The observation file grows to 3.8 MB and no score moves. 2. momentum becomes a list, one entry per span, shortest first. Ten years and twenty years are published. primaryMomentum returns the first entry for surfaces that show one number. 3. Every indicator result carries its own normalised series, one point per observed year, back to whatever the data supports.

Why. D22's matched basket is held to the shallowest indicator in a dimension, so one span had to choose between breadth and reach. Measured on Brazil, the basket falls from four indicators at ten years to two at twenty in Anticipation and Building, while Adaptability holds four all the way to twenty-five. Publishing both spans lets each dimension say how far its own evidence reaches instead of being cut to the shortest common span.

An indicator, unlike a dimension, is comparable with itself. Nothing has to be matched, so its line runs as far back as the series does: 36 points for Brazil on several World Bank series, against six for the Doing Business rows.

The twenty-year view already says something the ten-year view cannot. Building falls by a median of 5.9 points across all sixteen countries over twenty years and is flat over ten. Colombia is at −14.3, Argentina −11.5, India −10.4, Brazil −4.7, and Korea is the one clear gain at +11.3.

Cost.

• scores.json grows to 2.1 MB because every indicator now carries its history. The viewer reads it per request, which is fine locally and would need an API if this were ever served at scale.
• Two numbers per cell invite cherry-picking the flattering span. Both are always printed together with their basket sizes.
• Reaching back to 1990 covers a period when several of these countries changed political and economic system, so a twenty-year line crosses a discontinuity the model cannot see.
• Nothing is interpolated or extrapolated. A gap in a line is a real gap, and a series that stops, such as every Doing Business row in 2019, simply stops.

Overturned by. Enough indicator history to compute a full-dimension basket at both ends, which would make the matched basket unnecessary and both spans directly comparable to the headline score.

D25: Every point carries its provenance, and every run records what moved

Recorded 2026-08-26.

Choice. Two changes, both aimed at making the history checkable before the country set grows.

1. Each point in an indicator's series carries the value as published, the normalised value and its own source tier. A chart can be inspected point by point rather than trusted. 2. Every ingest compares itself against the file it is about to replace and appends what moved to data/observations/revisions.json: values restated, years added, years dropped, with the before and after for each. A full copy of the observation file is written to data/observations/snapshots only when --snapshot is passed.

Why. The audit trail claimed more than it delivered. Provenance was complete at the file level and absent from the rendered line: the viewer drew a series of normalised numbers with no raw value, no tier and no way to check any of it.

The second half matters more. A published statistic is not fixed. Agencies restate, rebase and revise, and an ingest that overwrites its own file makes that invisible, so the record would always claim a number had been what it is now. That is the failure mode a benchmark cannot have.

Per-point tiers exist for what comes next. A series will mix an international republisher with a national statistics office as soon as national sources are added, and the reader has to see which point came from where. The World Bank is a republisher of IBGE, MCTI and the rest, so today every point says international_organization and that is itself worth showing.

Cost.

• scores.json grows from 2.1 MB to 3.0 MB. The viewer reads it per request. At 60 countries this file has to become per-country files or an API, and that is the next structural limit rather than a future one.
• A full snapshot is 3.8 MB, so snapshots are opt-in and gitignored, and the revision log is the record that ships with the repository. The log stays small because it holds only what changed. Anyone wanting bit-for-bit archives of every run has to keep them outside git.
• A run that re-baselines everything would write one very large entry, so the list is capped at 500 revisions per run with the remainder counted in omitted.
• The log starts now. Everything ingested before today has no revision history and never will.

Overturned by. A move to per-country output files, which would change where the series lives but not what it has to carry.

D26: Every term is defined once, in plain language, in the model

Recorded 2026-08-26.

Choice. packages/core/src/model/glossary.ts holds a definition of every term this project invents or borrows: dimension, indicator, measurement class, score, reference frame, normalisation, winsorizing, out of frame, confidence and its three parts, confidence band, source tier, gap, retired indicator, evidence record, momentum, matched basket, indicator line, Delphi panel, provenance, dissent, wealth proxy and known artefact. Each entry carries a one-line version, a full explanation written for somebody who has never seen the benchmark, and a worked example from the current data.

The viewer renders it at /glossary. Measurement classes get their own block because they appear as a bare letter everywhere else. ClassLegend ships under every table that shows those letters, and country pages open with a short guide to reading the page.

Why. The viewer was showing a reader the letter C, a dashed radar edge and a confidence band of 0.079 and assuming all three were self-explanatory. None of them are. A benchmark that wants to be argued with has to be legible first, and a definition sitting only in DECISIONS.md is not available to the person looking at the chart.

Putting the glossary in the model rather than in the page matters for the same reason the confidence thresholds live in one file: two explanations of one term drift, and the drift is invisible until somebody quotes the wrong one.

Cost.

• Definitions age with the model. An entry naming a current number, such as Coordination sitting at 0.079 confidence, is stale the moment that changes. Examples are marked as examples for that reason, and a model change now has to update its entry in the same commit.
• The glossary is prose in a package that is otherwise data and arithmetic.

Overturned by. Nothing foreseeable. A second surface that needs the same definitions, such as a printed report or an API, reads the same file.

D27: Forty countries, and output split one file per country

Recorded 2026-08-26.

Choice. Twenty-four countries join the extended set, taking the benchmark to 40: Germany, France, the United Kingdom, Spain, Poland, Sweden, Finland, Ireland, Canada, Australia, Japan, China, Indonesia, Vietnam, the Philippines, Malaysia, Thailand, Turkey, Israel, the United Arab Emirates, Nigeria, Kenya, Rwanda and Ethiopia. The reference ten are untouched.

Scoring now writes data/out/index.json, the slim list of nine scores per country, and data/out/countries/{ISO3}.json, one country in full. The single scores.json is gone. It had reached 7.3 MB and the viewer read all of it to draw any page.

Why, on the countries. Every correlation in the diagnostics was a hint on 16 points. At 40 the picture changes and some of it reverses:

The headline is the duplicate list. At 16 countries, Anticipation and Coordination correlated at 0.94 and Anticipation and Agency at 0.94, which suggested the nine dimensions were three signals wearing nine names. At 40 countries, no dimension pair passes the redundancy threshold at all. The dimensions separate once the country set is wide enough to separate them, and the earlier finding was substantially an artefact of a narrow, mostly rich sample.

Why, on the files. A country page needs one country. A grid of 40 radars needs nine numbers each. Serving 7.3 MB for either is the scaling limit D25 predicted, arriving exactly where it said it would.

Cost.

• Clamping is now common, and that is a warning. 165 of 1,303 observed cells sit outside the reference frame and are clamped, 12.7 percent, concentrated in Ethiopia, Nigeria, Rwanda and Kenya. A10 said watch this flag, and it is now firing loudly. The ten reference countries do not span the range being asked about, and a low-income country pinned at 0 on twenty cells is losing real information. A versioned rebase of the frame is now a live question rather than a hypothetical one.
• The GEM values cover only the original 16 countries, so the two Experimentation indicators are missing for the 24 new ones and their coverage is lower.
• Twenty-four countries arrived with no Delphi estimates, so the panel now covers 16 of 40.
• The observation file is 9.2 MB.

Overturned by. A frame rebase, which would be a versioned event with its own decision, or a move to absolute anchoring per indicator.

D28: Icons are copied in, one per concept, never alone

Recorded 2026-08-26.

Choice. apps/web/src/components/Icon.tsx holds the path data for 23 Lucide icons, copied from lucide.dev rather than installed as a package, and credited in NOTICE.md. Concept-to-icon maps live beside them: measurement class, row status, source tier and glossary group.

Three rules govern their use.

1. One icon per concept, reused everywhere that concept appears, so the glyph becomes learnable instead of decorative. 2. An icon never appears alone. The measurement class badge still prints its letter, the trend still prints its number and sign, and the status cell still prints its word. 3. Every icon is aria-hidden, because the text beside it is the accessible name.

Why. The viewer asked a reader to decode a bare letter, a dashed line and a band. D26 fixed the words. An icon carried alongside those words gives the eye something to recognise at a glance in a dense table, which is where most of this data is read.

Copying the paths rather than installing lucide-react keeps the web app on three dependencies, which is a standing choice here, and it means an icon cannot change under us on a package update. The cost is that updating an icon is manual, which is the right trade for a set this small.

Cost. A reader who does not recognise a glyph loses nothing, because rule 2 holds, but a reader who misreads one could be briefly misled. Icons were chosen so that a wrong guess stays close: a target for the direct measure, a plug for an input, a rising line for an outcome, an eye for a perception.

Overturned by. A need for many more icons, at which point installing the package beats maintaining a copied set.

Update, same day. Extended to 36 icons. The nine dimensions each have one, used on the country page headings, the dimension table and the method page, so Building is always a hammer and Trust is always a handshake. The four confidence bands get a signal-strength ramp beside the meter, which is a third encoding of the same quantity alongside the bar length and the printed number.

D29: The eighth dimension keeps one name, and the axes carry their marks

Recorded 2026-08-26.

Choice. Two display changes, no change to any number.

The eighth dimension is labelled Building. The spec called it "Building / Execution" and the slash was doing no work: it read as two names for one thing and it broke every table column it appeared in. The dimension id stays building, because it is the key in every scored file, every trend basket and every Delphi estimate on disk. Renaming the key would invalidate all of it to change a word on a screen.

The label was briefly Execution on the same day. That word makes a promise the data cannot keep. Execution means finishing: cost performance, schedule performance, delivery. Those are large_project_delivery and firm_scale_up_rate, and both are gaps. What the dimension actually scores is four outcome measures of industrial output, manufacturing value added, high-technology export share, output per worker and economic fitness, plus a permit-speed score frozen at 2019. Calling that Execution invites a reader to take 9.4 as "Brazil cannot deliver" on the same page that documents Pix, SUS, the electoral system and Embrapa. Building is looser in a way that happens to be honest, and it matches Construção in the strategy this benchmark serves.

The radar takes a labels prop. full prints the dimension mark and the words, icons prints the mark alone, none prints neither. The 40-country grid uses icons, because at that card size the words rendered at seven pixels and were decoration. Every radar now also carries a <title> and a <desc> listing each dimension and its score, so a screen reader gets the full profile in words whatever the visual labels do.

Why. A reader learning nine glyphs once can then read 40 cards at a glance, which is the whole reason the grid exists. The written names stay on the large radar, in the dimension table, on the method page and in the accessible description, so the icons are never the only carrier.

Update, same day. Two corrections after seeing it in use. The axis marks are drawn at one strength for every dimension: fading them for thin evidence made the whole ring look washed out on a country where most dimensions are thin, and that evidence is already carried by the dashed edge, the hollow vertex and the asterisk. The radars are also drawn larger. The focal column is wider, the country grid went from four columns to three, and the geometry now depends on what labels the axes: words need a wide margin and shrink the shape to a small figure in a large box, marks need very little. The icon-labelled radar draws its shape across 78 percent of its box against 54 percent before, so at a 1280 pixel viewport the drawn spider on a country card went from about 135 pixels across to 256, and the labelled one from 205 to 287.

Cost. A reader who has not learned the marks has to visit a labelled radar first. The nine icons are a judgement: a telescope for Anticipation and a handshake for Trust are readable, while a shuffle for Adaptability and a hand for Agency are weaker. Those two are the ones to revisit if anybody misreads them.

Overturned by. Evidence that the grid is unreadable without words, which would mean going back to labels and making the cards larger.

D30: Every number opens onto the field it sits in

Recorded 2026-08-26.

Choice. Clicking an indicator name, an indicator's normalized score or a dimension score opens a panel listing every country on that measure, ranked, with the country being read marked and the ten frame countries marked.

Scoring writes a third artefact for this, data/out/indicators/{id}.json, which is the scored matrix turned inside out: one file per indicator holding every country's raw value, year, source tier, normalized position and whether it clamped. Two route handlers serve it and the dimension equivalent, and the panels fetch on demand rather than shipping 34 payloads with every page.

Why. A table cell reading 17.6 is not information. The same cell next to the other 39 countries, the two values that fix the ends of the scale and the year each country's number comes from is information. This is what Our World in Data does well and what a static table cannot do at all.

The panels also surface problems that the country page hides by construction. Opening R&D expenditure shows Israel at 6.35 percent of GDP and South Korea at 4.94 both sitting at exactly 100, because Israel is outside the frame and clamped, which is artefact A10 made visible instead of documented. Opening Coordination shows the Netherlands, France and Spain tied at 100 with confidence 0.079, which is A12 in one glance.

Cost.

• Another 392 KB of generated output, and another artefact to keep in step with the scores. It is written by the same command, so it cannot drift.
• The panels are client components with a fetch, so the context is not available without JavaScript. The raw value, year, source and normalized score all stay in the table itself, so nothing is lost, only the comparison.
• A dialog inherits text alignment from wherever it sits in the DOM. Both panels set their own, which is a small trap worth remembering for the next one.

Update, same day. Both panels now open with a distribution plot and keep the ranked list below it. Every country is a dot on the same 0 to 100 axis, dots that would overlap stack upward so a cluster reads as a column, and the box behind them is the middle half of the field with the median as a line. The country being read is filled and labelled. In the dimension panel a hollow dot means thin evidence, the same convention the radar uses.

The shape is the point. R&D expenditure puts a dozen countries in a single column at the floor of the scale and spreads the rest thinly across the top half, which a rank cannot show and a bar chart of 40 rows buries.

Overturned by. Nothing foreseeable for the panel itself. A reader who wants the underlying distribution rather than the normalized one needs a raw axis with a log option, which is a further step.

D31: A record carries its mechanism, and patterns get their own page

Recorded 2026-08-26. Extends D20.

Choice. An evidence record gains an optional pattern: the mechanism in one or two sentences, the preconditions that had to already exist, and where the move has travelled. All 15 records now carry one, and pnpm bench validate warns when a record does not.

The layer also stops being something a reader finds by opening a country. There is a /patterns page listing every record across every country, grouped by the dimension it bears on.

Why. The record already said what a country did. It did not say how, and how is the only part that transfers. Pix as a number is 7.98 billion transactions a month, which tells a Brazilian nothing they can act on and tells anybody else nothing at all. Pix as a mechanism is a central bank that wrote the standard, compelled participation above a size threshold, ran settlement itself and priced it at zero, which is a move somebody else can consider, and whose three preconditions tell them whether it would work where they are.

Everything else in a record is sourced from a named publisher. The mechanism is not: it is our reading. It sits in its own field for that reason, so the sourced and the interpreted are never confused.

Cost.

• The mechanism is an argument and can be wrong in a way a published number cannot. It is signed by being ours, and a reader who disagrees is disagreeing with an analysis rather than with a statistic.
• Preconditions are the weakest part. They are what makes a copy fail, they are the hardest thing to establish, and three lines per record is a first pass.
• Fifteen records across four countries is a library, not a discipline. The interesting version compares mechanisms across countries facing the same constraint, and that needs many more records.

Overturned by. Enough records to compare mechanisms rather than list them, at which point the page becomes a query rather than a list.

D32: Evidence is drawn as a gradient, and every chart is a control

Recorded 2026-08-26.

Choice. Four changes to how charts behave, and one naming change.

The radar edge no longer switches between solid and dashed at a threshold. Each edge is cut into fourteen segments, confidence is interpolated along it, and the gap between dashes opens as the evidence thins. An edge running from a well-evidenced dimension to a poorly evidenced one comes apart gradually, which is what the underlying quantity actually does.

The asterisk after a thin axis label is gone. The dashed edge and the hollow vertex already said it twice.

Axis labels are controls. Clicking one opens the 40-country panel for that dimension, the same panel a score opens.

Distribution dots respond to a pointer: the country under it is named where it sits, with its raw value, and the hit target is twice the size of the dot.

The product is called NCB in the wordmark, with the full name beside it.

Why. A benchmark whose whole argument is "the evidence is uneven" should draw the unevenness rather than annotate it. A threshold hides the gradient it stands on: Adaptability at 0.47 and Learning at 0.52 are nearly the same amount of evidence and were drawn as opposite states.

Making labels controls follows from D30. If a number opens onto its field, the name of the thing it measures should too, and a reader hunting for context should not have to learn which parts of a page are live.

Cost.

• Nine axes times fourteen segments times each series is a lot of line elements for one chart. It is still under a hundred and fifty and the radar has no dependencies, so it stays cheap.
• The gradient makes a single edge harder to read exactly. That is honest: the precise confidence numbers are in the table beside it and always were.
• Native dialogs centre themselves through a margin the CSS reset removes. Both panels set it back explicitly, which is the second trap this element has sprung after inheriting text alignment.

Overturned by. Nothing foreseeable. If the gradient reads as noise at small sizes, the icon-labelled radars can fall back to the threshold.

D33: Evidence records get an inclusion rule before they get more records

Recorded 2026-08-27. Extends D20 and D31.

Choice. docs/EVIDENCE.md states the inclusion rule for evidence records and how a record is written. Five tests decide whether a case gets in: it bears on a declared gap, a named publisher carries the number, the delivery is institutional, it was delivered rather than announced, and its limits can be written honestly. Three disciplines govern the corpus: at least one record in five documents a reversal, expansion goes by dimension rather than by country, and no country holds more than a third of the set. Records stranded by a gap promotion are deleted in the promoting change, with their ids named in its decision entry.

Why. D20 rejected curated national successes with no schema, no source discipline and no limits. The schema fixed source discipline and limits. It did not fix selection: nothing said who picks the cases, what disqualifies one, or how many failures the set must carry, and a validator cannot see that a corpus of pure successes is a brochure. Fifteen records exist, eleven of them Brazilian, twelve of them on one indicator, and six of nine dimensions have none. Expanding that to forty countries without a written rule reproduces the failure D20 exists to prevent, inside a schema.

The rule is written down now, before the corpus grows, because a rule adopted at fifteen records constrains the author and a rule adopted at a hundred and fifty indicts the archive.

Cost.

• The one-in-five reversal quota and the one-third country ceiling are round numbers, not derived ones. They exist to force the question at authoring time, not because five and three are correct.
• Deleting stranded records on promotion loses their rendered text from the site. The decision entry and git history keep it, but a reader of the patterns page does not see git history.
• The rule makes authoring slower. That is partly the point.

Overturned by. A corpus that satisfies every test and still reads as advocacy — that would mean selection bias lives somewhere the rule does not reach, and the rule needs to move from authoring discipline to independent review. Or a demonstrated need to document sub-national or non-state deliveries, which test three currently excludes.

D34: Countries are identified by ISO 3166-1 alpha-3, recorded after the fact

Recorded 2026-08-27.

Choice. The country identifier everywhere in this project is the ISO 3166-1 alpha-3 code: iso3 in countries.ts, the cell key in observations, the file name in data/out/countries/{ISO3}.json, and the route in the viewer. Alpha-2 is not used anywhere.

Why. The World Bank v2 API returns every observation row keyed by countryiso3code, and the request path takes the same code, which the pipeline stores verbatim as sourceUrl provenance. Alpha-3 is therefore the identity the raw data arrives with, and using alpha-2 would add a translation table between the source and the store for no benefit. The convention also matches the rest of the cross-country data world (IMF, ILOSTAT, Penn World Table), which matters because ILOSTAT is a planned adapter, and the codes read better as file names: CHE reads as Switzerland where CH invites confusion with China.

This entry records a constraint, not a choice that could have gone the other way. It exists because the question "why three letters" had no written answer.

Cost. None beyond the three characters.

Overturned by. A primary data source that keys on something else and outweighs the World Bank in the registry. That would justify an internal ID with per-source mappings, and this entry should be superseded when it happens.

D35: The agenda is computed, and language is an interpretation layer

Recorded 2026-08-27.

Choice. pnpm bench agenda turns each country's scored output into a capability agenda: a language-neutral JSON in data/out/agenda/{ISO3}.json and one rendered markdown per lexicon beside it. The generator classifies each dimension by two published thresholds. Confidence below the usable band makes a dimension a measure-first item, because the score cannot carry a decision. Usable confidence with a score under 50 makes it a raise item. Raise items name the three highest-scoring countries whose own evidence is usable, and the evidence records other countries filed against the dimension's gaps. The declared gaps across all dimensions form the measurement agenda. The subject country's own evidence records close the document, outside the numbers, as always.

Language lives in packages/core/src/i18n as lexicons: data files mapping the model's vocabulary and the agenda strings into one language each. The ground layer stays English end to end: ids, registry definitions, JSON output. A lexicon lookup that misses falls back to the registry English, so a partial lexicon renders complete pages. pt-BR is the first lexicon and the template for the next one.

Why. A hand-written national to-do list is advocacy the moment it is signed, and it goes stale the first time the data moves. A computed agenda is neither: it regenerates with every run, every claim in it traces to a score, a gap or a record, and a reader who distrusts a translated page can diff it against the JSON it renders. Separating lexicon from renderer makes a new language a data contribution rather than a code change, which is the shape of contribution the project wants most.

Cost. Two thresholds are now product decisions: the usable band already lives in confidence.ts, and the raise cutoff of 50 is a constant in agenda.ts with no empirical basis yet. Evidence record titles and claims render untranslated inside non-English documents, because records are ground data. Exemplar selection rewards measured countries: a country with real capability and thin evidence cannot appear as an exemplar, which repeats the benchmark's general bias toward the measurable.

Overturned by. A reader study or field use showing the raise cutoff misleads at 50, which would justify deriving it from the score distribution instead. Records gaining translated fields, which would remove the mixed-language cost. A lexicon whose translation drifts from the registry meaning, which would justify review rules for lexicon changes rather than plain PRs.

D36: A record states where the delivery stands, and a claim can carry two numbers

Recorded 2026-08-27. Extends D20, D31 and D33.

Choice. Two additions to EvidenceRecord.

status, required: where the delivery stands as of the record's retrieval date. One of operating, concluded, eroded, dismantled. A reversal is a record whose status is eroded or dismantled, tested by isReversal from @ncb/core, and pnpm bench validate warns when reversals fall below the one-in-five quota D33 set.

secondMetric, optional: a second published number in the same shape as metric, for the claims one number cannot hold. An eroded record pairs its current value with the peak it fell from, and the validator warns when it does not. A delivery record can pair scale with a cost or schedule figure.

All fifteen records now carry a status: twelve operating, two concluded (Plano Real, Luz para Todos), one eroded (the immunisation programme, whose peak of 99 percent in 2003 is now its second metric). Both surfaces that render records print the status and the second number.

Why. D33 requires the corpus to carry its reversals and gave the quota to the author's discipline, because free text is not countable. That made the rule's own enforcement section admit the validator was blind to the thing the rule most cares about. The immunisation record proved the point: its erosion, the reason it is in the set, lived entirely in limits prose and rendered as "running since 1973" — indistinguishable from Pix.

The second slot exists because the flagship indicator asks for cost and schedule performance and every record filed against it admits in limits that one scale number cannot show that. One optional slot is the smallest change that lets a record hold a claim with two sides.

Cost.

• Status is a judgement stamped at retrieval time and it goes stale silently. A programme dismantled after the record was written still reads operating until someone edits the file. retrievedAt bounds the staleness; nothing detects it.
• Four states force real cases into coarse bins. Proálcool nearly collapsed in the 1990s and is operating; the near-collapse stays in limits, and the bin says nothing about the path.
• Two metric slots invite cramming. The rule stays one delivery, one record; the second slot is for the other side of the same claim, not a second claim.

Overturned by. Records that repeatedly need a third number, or statuses that keep landing in the wrong bin — either means the lifecycle deserves a dated series, not two stamps, and the slot design should be replaced rather than extended.

D37: The output directory describes itself

Recorded 2026-08-27. Extends D27 and D30; D34 records the country identifier.

Choice. data/out becomes a self-describing dataset, built on standards a consumer's tooling already speaks.

1. JSON Schema. bench score emits a schema for each published shape into data/out/schema/: the index file, the country file and the indicator view. They are generated from the Zod schemas in packages/core/src/model/schema.ts by zod-to-json-schema, so the Zod definitions stay the single source of truth and the emitted schemas cannot drift from what the pipeline writes. This is the package's one new dependency, taken so the schemas would not be hand-written copies. 2. Data Package. data/out/datapackage.json is a Frictionless Data Package descriptor naming every published file, its schema, its source and its license. Standard data tooling can consume the directory from that one file. 3. Semantic versioning. The dataset carries a version, defined once in packages/core/src/model/version.ts and stamped into the index, every country file and the descriptor. Major = the frame rebased or a published field removed, which is the "versioned, announced act" the frame invariant already required without naming a scheme. Minor = countries, indicators or fields added. Patch = a re-ingest under the same registry. The version describes the contract, not the method; KNOWN-ARTEFACTS.md tracks the method. First stamped version: 1.0.0. 4. Data license. The derived dataset is CC BY 4.0, matching the World Bank data it derives from. Stated in NOTICE.md and in the descriptor. 5. RFC 4180. table.csv now uses CRLF line endings and quotes fields containing CR, so it is a conforming file rather than a nearly conforming one. 6. Schema.org. The viewer's landing page embeds a JSON-LD Dataset block, which is what dataset search engines index.

Why. The project's boundary rule is stable IDs and standard formats, because the output is meant to be consumed by things this repository does not know about. Until now that contract lived only in TypeScript types, which a non-TypeScript consumer cannot read, and the version of what they were reading was not written anywhere. Each of the six is the smallest standard that closes one of those gaps. Full SDMX and RDF were considered and rejected as institutional-publisher machinery; keeping source series codes verbatim as provenance borrows the useful part.

Cost.

• One new dependency in @ncb/core (zod-to-json-schema), against a standing preference for zero. The alternative was a second, hand-maintained copy of every published shape, which is the drift this repository's DRY rule exists to prevent.
• The version is bumped by hand. A forgotten bump mislabels a release; the bump rules sit on the constant to make that harder.
• CRLF in table.csv will show as a whole-file diff against the previous LF file exactly once.

Overturned by. A consumer that standard tooling cannot serve from the Data Package, which would argue for a real API. Or the schemas drifting from the files in practice, which would mean generation is wired to the wrong place and validation of the emitted output should be added to bench validate.

D38: The homepage is global, and one country is a layer on top of it

Recorded 2026-08-27. Supersedes the focal-case layout noted in D29.

Choice. The homepage no longer leads with one country. It opens on the grid of every country's shape, sorted alphabetically, followed by the score table and the confidence table. FOCUS_ISO3 is deleted from apps/web/src/lib/profile.ts, and the page title asks about a country rather than this country. CompareRadar stays where it belongs: on a country page, where the reader has already chosen a subject.

The dashed-edge legend and the frame note now ship under the grid. Both were previously reachable only inside CompareRadar, so a reader who never opened a country page never learnt what a dashed edge meant. RadarEvidenceLegend takes an interactive prop, because the grid radars pass no onSelectDimension and the legend was promising a click that is not there.

The method page loses its Brazil bullet too. What that bullet actually argued is that a frame fitted to one country describes that country, and it makes that argument without naming one.

A country-specific entry point is a layer above this page, not the spine of it. It is not built yet.

Why. Two readers arrive here. One works inside a country and wants that country. The other wants to know what the benchmark measures and whether the method holds. Leading on Brazil served the first and asked the second to read past a case they did not choose. The grid answers both: it shows nine dimensions producing 40 different shapes, which is the claim the project actually makes, and every card is a door into a country. The measurement is global, so the front door should be too.

Cost. The homepage loses its one large, readable radar and the compare control. A first-time reader now meets the nine axes at icon size and has to open a country to see the shape drawn with words on it. DimensionLegend carries the names on the grid, and the accessible description in every radar carries them in full, but that is a legend, not a labelled chart. If the icons turn out to need a worked example first, the fix is a single labelled radar above the grid, not a return to a focal country.

Overturned by. Use showing readers cannot enter the grid without a worked example. A decision to make this viewer a country-specific product, which would put the focal case back and make the global grid the secondary surface.

D40: A date is metadata, and a document a reader is told to open is a link

Recorded 2026-08-27. Extends D35.

Choice. Two presentation rules, both applied everywhere a document is rendered.

A generation date is metadata about the document, so it sits under the title on its own line and never inside a sentence. The agenda lexicons carry it as agenda.generated, separate from agenda.intro, and the renderer prints it as a dateline: *Gerado em 2026-08-27 a partir da rodada de dados atual*. The viewer prints the same string in the muted metadata size under the page title. report.md splits its dateline off the same way. The intro that follows now opens on what a score means, which is the first thing a reader needs.

A repository file a rendered document tells a reader to open is a link. The URL is built by docHref in packages/core/src/model/project.ts, which also holds REPO_URL, LIMITS_DOC and DECISIONS_DOC. The agenda intro links docs/KNOWN-ARTEFACTS.md, the report links docs/DECISIONS.md, and datapackage.json reads its homepage from the same constant instead of repeating the URL.

Why. "Gerado em 2026-08-27 a partir da rodada de dados atual." read as the first claim the agenda makes, ahead of the claim about what the numbers are. Nobody reads a document for its build date.

The linking rule fixes a worse problem. The agenda tells every reader to read the limits before quoting a number, and it named a path only somebody with a checkout could open. The instruction was correct and the reader could not follow it. A published document reaches people who will never clone the repository.

Cost.

• docHref pins the main branch, so a link from an old rendered document points at the current file rather than the file that was current when the document was written. The alternative, a commit SHA, would make the output churn on every run.
• The lexicons carry one more key, and a partial lexicon that omits generated falls back to English while the rest of the page is translated.
• Markdown emphasis is the whole dateline treatment. A renderer that ignores emphasis prints the date as an ordinary line.

Overturned by. A viewer page that renders the limits document itself, which would give the viewer a local target for {limits} while the markdown keeps the repository link. Or a versioned documentation site, which would replace the branch in docHref with a release.

D41: The pipeline stops discarding its own warnings, and the viewer carries them

Recorded 2026-08-27. Extends D40 and takes up its overturn clause; extends D12, D23, D25, D35. Prompted by a full coherence review of the model and the viewer.

Choice. A set of repairs with one principle: a warning the pipeline computes is published and rendered, and a claim a surface makes is derived from the data it sits above, never hard-coded beside it.

• Momentum.clamped was computed and read by nothing. It now travels: AgendaTrend carries it, the report marks clamped cells, the agenda documents and the viewer print it beside every trend, and the glossary example for momentum names it. Brazil's headline "+26.2 on Agency over ten years" carries "2 of 4 clamped" everywhere it appears, because part of that delta is boundary distance rather than movement.
• Clamping is aggregated. Diagnostics.outOfFrame counts clamped cells overall and per country, the report prints it, and the diagnostics page renders it. 12.7% of observed cells clamp in the current run, concentrated in the lowest-income extended countries, which is A10 stated as one number.
• Exemplars exclude clamped countries. An agenda offers its exemplars as somebody to learn from, and a clamped 100 is partly an artefact of the frame.
• evidenceElsewhere matches its contract: records filed against the dimension's declared gaps only, not any indicator in the dimension.
• dataGaps carries a status, so retired indicators stop being reported as "no dataset exists". A rejected dataset and a missing one are different claims, per D23.
• The report derives its confidence headline instead of asserting "no pair reaches the good band" (0.67 crossed the threshold and the sentence shipped false). Diagnostics page headings are computed from the tables under them for the same reason: two of them had gone stale enough to be wrong.
• The dissent threshold has one home, DISSENT_IQR in the model layer. isPanel is now actually called; the viewer gates every Delphi surface on isEvidential and presents a non-panel run as "session estimate", never as "panel median". A mock run renders nowhere as evidence.
• A failed World Bank fetch carries the previous file's observations forward for that series instead of dropping them, so a transient error can no longer be written into the revision log as the publisher removing decades of data.
• The viewer renders docs/KNOWN-ARTEFACTS.md at /limits, in the nav, with one anchor per artefact. The agenda's {limits} points there in the viewer and stays on the repository link in the rendered markdown, which is the split D40's overturn clause anticipated and MZ asked for.
• Language switching left the nav. It is one control in the layout header, driven by languageCounterpart in apps/web/src/lib/links.ts, shown only where a counterpart page exists. Every URL shape lives in that file again.

Dataset version goes to 1.1.0: fields were added, nothing published moved.

Why. The review found the same failure four ways: comments promising what no code enforced, warnings computed and thrown away, findings hard-coded above the data that had moved past them, and caveats that stopped at the repository boundary while the numbers crossed it. Each repair is small; the pattern was the risk.

Cost.

• Computed headings can read awkwardly ("None of nine dimensions..."), and a heading that is always true is less quotable than a sharp claim.
• Excluding clamped countries from exemplars can leave a dimension with fewer than three exemplars even when well-scored countries exist.
• The carried-forward observations of a failed series age silently until the next successful fetch; only the ingest report on stdout says it happened.
• /limits renders an internal document written in British English with em dashes, against the viewer's copy rules. The renderer converts the heading dashes and leaves the body verbatim, because mangling the record would be worse than the style breach.

Overturned by. A run where the exemplar exclusion empties most dimensions, which would argue for annotating clamped exemplars instead of excluding them. A persistent ingest failure record, which would replace the silent carry-forward.

D42: A candidate is judged on what it does to its dimension, not on its own correlation

Recorded 2026-08-27. Extends D23; answers A12.

Choice. bench diagnose now emits wealthAttribution: for every indicator, its dimension's correlation with log GDP per capita as published, the same correlation with that indicator dropped from the mean, and the difference. The diagnostics page prints it, sorted by the indicator that raises its dimension's wealth correlation most. A candidate indicator is accepted or rejected on that delta, and no longer on its own correlation alone.

The counterfactual is computed from the matrix rather than read from the published score, so dropping a row means recomputing the dimension mean exactly the way score.ts does, with missing values dropped and nothing imputed.

Why. The existing test asks whether one series tracks income and flags it above 0.70. That is not the question the benchmark's central claim rests on. A dimension can hold indicators that each sit under the line and still track income as a group, and an indicator under the line can still make its dimension worse.

This was found by making the mistake. A probe of 25 candidate World Bank series, run to answer A12, produced one usable observable trust measure: IC.FRM.CORR.ZS, the share of firms expected to give gifts to public officials. It resolves for all ten reference countries, it is behavioural, and it correlates with log GDP at 0.667, under the line. It was wired, ingested and scored. It raised Trust confidence for 34 of 40 countries from very thin to thin, and it moved Trust's own correlation with GDP per capita from 0.385 to 0.619. It was reverted the same session, and 0 of 360 published cells now differ from before the attempt.

The first run of the new diagnostic then found something the project did not know. homicide_rate raises Trust's wealth correlation by 0.288, from 0.096 to 0.385. It is the largest single wealth contribution in the model, and it is the indicator D23 added as the observable replacement for the retired perception composites. contract_enforcement_days runs the other way at -0.189: without it Trust would correlate at 0.573. The dimension the project treats as its most income-contaminated is contaminated by one row, and that row was added to fix contamination.

Cost.

• The delta is computed against the current country set and moves when countries are added. It is a diagnostic and not a threshold, and no rule fires on it.
• It cannot separate an indicator that imports wealth from an indicator that correctly measures a capability wealthy countries genuinely have. Homicide may be either. The number says where to argue, and it does not settle the argument.
• Nine dimensions times 34 indicators means the dimension mean is recomputed 34 times per run. It is not measurable next to ingestion.

Overturned by. Evidence that a high delta is routinely the right answer, which would make the diagnostic noise. Or a dimension-level wealth test derived from a partial correlation rather than a leave-one-out, which would be the stronger statistic if the country set ever grows enough to support it.

D43: The country page opens on the agenda, and the split has one home

Recorded 2026-08-27. Completes D38; extends D35. Originally written as D39, which a concurrent session overwrote; D39 is deliberately left unused.

Choice. /country/{ISO3} opens with a lede computed from data/out/agenda/{ISO3}.json: the strongest dimension where the evidence is usable, the first dimension the evidence says to raise, the first that cannot be judged at all, then two lists, raise items lowest score first and measure items thinnest evidence first. Every entry links to that dimension's section further down the same page, and the block links out to the full agenda.

The lede selects and never calculates. Every score and confidence it prints comes straight out of the agenda JSON. The sorting comes from splitAgenda in packages/core/src/pipeline/agenda.ts, which AgendaView calls as well, so one function decides which dimension leads and the country page and the agenda document cannot disagree.

The eyebrow names the country's role in the frame, and the registry reason moves to a quiet line under the agenda. RAISE_BELOW is exported, because the page states the threshold when a country has no dimension above it.

This is the layer D38 said was not built. It is not specific to one country: any country with an agenda file gets it, and Brazil is only the country somebody asked about first.

Why. A reader landing on a country met a radar and nine tables, and had to read all nine to learn where to look. The agenda already computed that answer for the markdown documents and the viewer was not using it. Putting it at the top costs no new computation and opens the page on a finding. Brazil reads "Nothing here scores above 50 on evidence strong enough to act on", which is the honest headline for that country and was previously buried nine sections deep.

Cost.

• The lede repeats what the sections below say. A reader who scrolls meets each named dimension twice.
• hold is a real agenda kind with no list of its own here. Only the top one appears, in the first sentence, so a country with eight holds shows one.
• The threshold sentence prints RAISE_BELOW as a bare number, so a reader can meet 50 before meeting the frame note that says what 0 to 100 means.

Overturned by. A country page that reads better with the shape first and the agenda second, which would move the block under the radar. Or a hold list long enough to deserve a column of its own.

D44: The homicide rate is retired, because it was the wealth signal it was hired to remove

Recorded 2026-08-27. Supersedes the Trust half of D23. Evidence from D42.

Choice. homicide_rate becomes ingest: 'retired'. The row stays, is never fetched or scored, and lowers confidence exactly as a gap does.

Why. D23 added it as the observable replacement for two retired WGI perception composites, on the argument that it is counted by police and health systems rather than reported as an opinion. D42's leave-one-out diagnostic then measured what it actually did to the dimension. It raised Trust's correlation with log GDP per capita by 0.288, from 0.096 to 0.385, the largest single wealth contribution anywhere in the model. Removing it moves Trust to 0.097 and changes no other dimension.

The objection is not to the dataset, which is sound, and not to the reasoning in D23, which was right about the difference between an outcome and an opinion. It is that across this country set the variation homicide carries is mostly income. Homicide is driven heavily by organised crime, and a society can be physically safe while trusting very little. The benchmark exists to show that capability is a separate property from wealth, and this row was quietly arguing the opposite.

Cost. Trust now has one observed indicator of seven, contract_enforcement _days, frozen at 2019, and its confidence falls from 0.191 to 0.079. The dimension is now honestly unmeasured where it was previously measured wrongly. That is a worse-looking model and a truer one, and it forces D45.

Overturned by. A country set wide enough that homicide stops tracking income, which would mean the correlation was an artefact of ten reference countries. Or a behavioural trust measure landing beside it, which would let the dimension carry homicide without homicide carrying the dimension.

D45: A dimension with fewer than two observed indicators publishes no score

Recorded 2026-08-27. Answers the question A12 left open. Forced by D44.

Choice. MIN_INDICATORS_FOR_SCORE is two, in packages/core/src/pipeline/score.ts. Below it DimensionResult.score is null, belowCoverageFloor is true, and observedIndicators says how many there were. Two is the same minimum D20 requires before a gap is promoted to a scored indicator, so the model uses one number for the same idea.

Confidence, the indicator rows, the evidence records and the trend all still publish. What is withheld is only the average.

Three surfaces changed with it. The radar leaves an unmeasured axis empty and closes the shape across the gap, where before it plotted the missing value at the centre and drew a country as catastrophically weak on a dimension nobody had measured. DimensionScore in the viewer prints "not measured" with the indicator count on hover, so the reason is available rather than implied. The flat table and the country tables read from it.

This withholds 84 of 360 published cells: Coordination and Trust for all 40 countries, plus Experimentation for one country and Shared Purpose for three.

Why. Coordination printed 46.7 for Brazil off a single border-time measure frozen at 2019, and moved from 15.5 to 46.7 without anything changing in Brazil. A12 recorded that as an artefact and mitigated it with a dashed edge and a confidence band, then said plainly that the mitigation is not a fix. A mean of one number is not a measurement of a dimension. Printing it invites exactly the decision the evidence cannot carry, and the display treatments were asking the reader to discount a number the model should not have offered.

Cost.

• Two of nine dimensions now show nothing for every country. A reader meeting the benchmark for the first time sees a seven-sided shape and has to learn why.
• blendedScore falls back to the panel when the floor withholds a score, which is a wider fallback than the invariant described. The panel guards still hold: a non-evidential run is never presented as evidence.
• The floor is a count and ignores what the indicators are. Two weak indicators pass and one strong one does not.
• Comparisons against anything published before today lose two dimensions.

The dataset goes to 2.0.0. D37 reserves major for a rebased frame or a removed field, and neither happened here. But a consumer parsing score as a number now gets null for 84 of 360 cells, which breaks them exactly as a removed field would. The rule is about what breaks a reader, so the version follows the break.

Overturned by. Replacement indicators landing in Coordination and Trust, which is what A12 asks for and would make the floor moot for them. Or evidence that readers treat an empty axis as a zero anyway, which would mean the withholding needs words on the chart and not only in the table.

D46: A case study is an address, and the list of them is a filter

Recorded 2026-08-27. Extends D20 and D33.

Choice. Every evidence record gets its own page at /patterns/{id}, and the /patterns index becomes a filtered view over the corpus.

evidenceHref in apps/web/src/lib/links.ts now writes /patterns/nld-delta-programme instead of /patterns#nld-delta-programme. The record id is the slug, so the address is already in the data and no second identifier exists. The record page carries the full record, the indicator it bears on, and two related lists: the other deliveries from the same country, and the same indicator in other countries.

The index keeps its dimension grouping and gains five controls: a text search over the title, claim, limits, mechanism, preconditions, country and publisher; a country select; a dimension select; a status select that also offers reversals as one class; and a switch for records that carry a mechanism.

The filters are the query string, ?q=, ?country=, ?dimension=, ?status= and ?mechanism=1. readPatternFilters and patternsHref in apps/web/src/lib/links.ts parse and build that shape, and both the server page and the client view use them, so the address is read where it is written. The page reads searchParams and renders the narrowed list, which means a shared link shows what the sender saw with no client round trip. Each control change rewrites the address with history.replaceState, so a keystroke never reaches the server. Unrecognised values are dropped and the full list renders.

One card renders in both places. apps/web/src/components/PatternCard.tsx holds the metadata line, the metrics line, the mechanism block and the limits line, and the index and the record page both read from it.

Why. The corpus reached 33 records across 19 countries, and an anchor into one long page is not a citable thing. A reader who wants to send somebody the Delta Programme record sends a whole page and a scroll position, the browser lands them mid-list with no context above the fold, and a search engine indexes one page for 33 deliveries. D33 asks the corpus to carry its reversals, which only works if a reversal can be pointed at.

The filters follow from the same growth. Grouping by dimension was enough at nine records. At 33 the reader has a question, usually about one country or one kind of loss, and scrolling is the wrong answer to it.

Cost.

• Any link written against the old anchor form now lands on the index without scrolling. Nothing outside the viewer wrote one, and evidenceHref was already the only place that built it.
• The index sends the whole corpus to the browser to filter it. At 33 records that is small, and it will not stay small. The filter state is in the URL already, so moving the filtering to the server is the next move and needs no new contract.
• replaceState keeps 30 filter changes out of the history stack, and the cost is that the back button does not step through them. It leaves the page.
• The query string is a published surface now. Renaming a dimension id or a status value breaks a link somebody saved.
• Two more routes to keep in outputFileTracingIncludes reasoning. Both read data/evidence, which is already listed.

Overturned by. The corpus growing past the point where shipping it to the browser is reasonable, which makes the index a server-rendered list reading the same query parameters. Or evidence that readers never use the record pages, which would mean the anchor was enough and the cost was the two extra clicks to reach a mechanism.

D47: Every country sets the frame it is scored against

Supersedes D16. Recorded 2026-08-27.

Choice. An indicator's Tukey fences and its 0 and 100 endpoints are computed over every country in the benchmark. There is no reference set and no extended set. The frame field is gone from packages/core/src/model/countries.ts, and REFERENCE_ISO3, EXTENDED_ISO3 and COUNTRY_FRAMES are gone with it. buildFrame takes all transformed values.

Stability moves from a privileged subset to the version number. The frame holds still inside a published dataset version. Adding a country rebases the frame, restates every score, and takes a major bump. That amends the D37 bump rules: a country addition is no longer minor.

Why. D16 bought comparability across runs by scoring 30 countries against a ruler built from the other ten. That is a measurement claim the data cannot support. A country measured against a frame it is absent from is not being placed among its peers; it is being told its distance from ten countries chosen in the prototype to expose contrasts. When the value falls outside their range the number stops being a measurement at all and becomes the sentence "beyond these ten", written as 0 or 100.

D27 already recorded the cost arriving: 165 of 1,303 observed cells clamped, 12.7 percent, concentrated in Ethiopia, Nigeria, Rwanda and Kenya. A10 said watch the flag. The flag was firing on one seventh of the low-income evidence. The stability D16 protected was real, but it was the stability of a scale that had stopped describing a third of the countries on it.

What the rebase did. At 40 countries, 356 of 360 dimension cells moved. The mean move is 3.8 points and the largest is 21.7 (South Africa, Agency). Clamped current cells go from 165 to 0, which is now structural: a value cannot fall outside a frame its own country helped build.

Cost.

• Every published number from 2.0.0 is restated. The dataset is 3.0.0 and the two are not comparable. Anything quoting a 2.0.0 score is stale.
• Every future country addition is a rebase. This is the guarantee D16 bought, given up on purpose. Adding a country is now announced, versioned and followed by a full rescore, and it can no longer be a side effect of loading data.
• Correlation with income rose. Anticipation 0.85 to 0.87, Agency 0.79 to 0.81, Building 0.56 to 0.59, Experimentation 0.40 to 0.52, Shared Purpose 0.33 to 0.46. Unclamping the low-income countries restored their spread, and that spread is largely income-ordered. The clamp had been hiding the strength of A3, not weakening it. No dimension pair passes the redundancy threshold, so D27's finding survives the rebase.
• outOfFrame now means something narrower. A current cell cannot set it. It fires on history scored against the current frame, where it still carries the trend caveat: 53 of 485 momentum baskets hold a clamped member.
• The frame is now sensitive to which countries happen to be loaded, which is what D2 was criticised for. The difference is that the sensitivity is declared, versioned and rescored rather than silent.

Overturned by. Absolute anchoring per indicator, which would remove the country set from the scale entirely and make a score comparable across versions. It needs a defensible floor and ceiling for each of the 33 scored indicators, hand-set and defended one at a time, and it trades a frame that describes this set for one that describes a claim about the world.

D48: The panel column takes the same wealth test as the indicators

Choice. diagnostics.json gains panelVsGdp. It correlates the published delphiScore column with log GDP per capita, dimension by dimension, and prints the indicator score for the countries the panel covered beside it. The WEALTH_CORRELATION_THRESHOLD of D42 flags a dimension the same way it flags an indicator. backfillCandidate marks the dimensions that publish no indicator score at all, because those are the only places a panel estimate could become the published number. A run that is not evidential reports its provenance and no rows: correlating mock estimates would produce a figure that reads as a finding.

Why. The panel was described as evidence the indicators cannot reach. It is not independent evidence. A language model reads the same published record the indicators are drawn from, plus press coverage, so the perception layer D23 retired can return through the panel wearing a new name. D42 tests that mechanism on indicators and nothing tested it on the panel. Every Delphi surface gated on provenance and panel size, which are questions about who produced the number, and none asked what the number tracks.

What it found, on the run active today. The in_session run, one panelist, 16 countries. Eight of nine dimensions of the panel column correlate with log GDP per capita at or above 0.70: Adaptability 0.93, Learning 0.92, Anticipation 0.91, Coordination 0.86, Building 0.84, Trust 0.83, Agency 0.80, Experimentation 0.77. Shared Purpose is the exception at 0.55. On Learning the panel is 0.15 above the indicators for the same countries, on Experimentation 0.14, on Shared Purpose 0.20.

Coordination and Trust are the finding. Neither publishes an indicator score under D45, so the panel is the only candidate for filling them, and the panel column on both sits above the line that retired the Worldwide Governance Indicators. Filling those two dimensions from this panel would restore the measurement D23 removed.

Cost. One correlation per dimension, and a section in the report and the diagnostics page that says the panel layer fails its own test. n is 16 and the panel is one analyst, so this is a hint under A8 and not a result. It is enough to stop the backfill and not enough to condemn a gateway panel that has never been run.

Overturned by. A gateway run whose panel column holds under 0.70 on the dimensions the indicators cannot measure. That is the evidence that would let a panel estimate carry a dimension, and this diagnostic is how it gets checked.

D49: The fetch describes itself once, and the viewer prints that description

Recorded 2026-08-28. Extends D37 and D40.

Choice. How a value reaches the dataset is declared in packages/core/src/model/sources.ts and nowhere else. The file holds the API base, the World Bank database ids with what a reader must know about each, the first year every series is asked for, the reader-facing label for each ingest route, and the request builder. pipeline/ingest.ts builds its calls from it. The new /sources page prints the same call and groups the registry by publisher from the same file.

Why. Provenance was published per indicator and nowhere in aggregate. The registry row carried a publisher and a link, the method page ranked source tiers by weight, and neither told a reader that 31 of 67 indicators come from one API, that eight World Bank series sit in the registry unscored, or which database a series needs. That last one is not decoration: the v2 API answers "indicator not found" for a code outside World Development Indicators when the request carries no source parameter, and until now that fact lived only in AGENTS.md, where no reader of the site can reach it.

The alternative was a hand-written sources page. A page that describes a fetch in prose drifts from the fetch on the first registry change, and the drift is silent because nothing compiles the prose. Declaring the shape once and generating the page keeps the two in step by construction, which is the same argument D40 makes for building document links from project.ts.

The data package descriptor also stops hand-listing its sources. It now names every publisher that supplies a value, from the same function, so a publisher that starts supplying one appears in the machine-readable descriptor without anybody remembering to add it.

Cost. One more file in the model layer, and a page that reads data/out/indicators/*.json to count what each publisher currently supplies. That is 39 small reads on one page, against the country files D27 forbids. WB_DATABASES and the trap table in AGENTS.md now say overlapping things and have to be kept in step by hand.

Overturned by. A second ingester. The World Bank shape is generalised in sources.ts only as far as one publisher needs, and an ILOSTAT or OECD adapter would want a per-publisher description rather than a World Bank one with other publishers listed beside it.

D50: The falsification conditions are an index, not a second document

Recorded 2026-08-28. Extends D40 and D49.

Choice. The site gets a /challenge page, and everything on it is derived. The open artefacts come from the severity line of each entry in docs/KNOWN-ARTEFACTS.md. The falsification conditions come from the Overturned by clause of each entry in docs/DECISIONS.md, newest first. apps/web/src/lib/docs.ts does the reading. Neither document gains a summary of itself, and the page holds no claim that is not already in one of them.

The footer carries the same idea in one line on every page: the dataset version, a link to the challenge page, the repository and the license split.

Why. Phase one of this project asks researchers to try to falsify the framework. Until now the site invited that in prose and offered no way to do it: no repository link outside a document body, no issue route, no citation string, and no version stamp a reader could quote with a number. The material for all of it already existed. Every decision has stated what would overturn it since D1, and CONTRIBUTING.md has stated the four contribution routes and the two-country rule for a gap. None of it was reachable from the viewer.

Hand-writing the list was the obvious alternative and the wrong one. A hand-written index of 48 falsification conditions is stale on the next appended decision, and the failure is silent. Deriving it means a new decision reaches the challenge page in the same commit that records it.

Cost. A markdown parser for a subset of a subset: the entry heading and two labelled paragraphs. It reads **Overturned by.** and **Severity:, so an entry that renames either label drops off the page without erroring. The docs/ writing convention is now load bearing in the viewer, which it was not before, and the two labels are checked nowhere.

Overturned by. A decision entry that needs more than one falsification condition, or an artefact whose severity is a judgment rather than a word. Both would mean the structure belongs in data rather than in prose, and the entries would move to a file the documents render from instead of the other way round.

D51: Latin America is covered whole, not sampled

Recorded 2026-08-28. Applies the D47 bump rules: dataset 4.0.0.

Choice. Every sovereign Latin American country is in the benchmark: Bolivia, Paraguay, Ecuador, Venezuela, Panama, Guatemala, Honduras, El Salvador, Nicaragua, the Dominican Republic, Cuba and Haiti join the eight already present, for 20 of 52 countries. LATAM_ISO3 in packages/core/src/model/countries.ts is the one definition of the region, and the Portuguese edition renders the whole set.

Why. The project's first field case is Brazil and its first institutional audience is Brazilian and regional (see docs/WHY.md). A regional comparison sampled at eight countries invited the question "why these eight" from exactly the audience the work is for. Covering the region whole removes the selection question, makes the benchmark legible to regional institutions, and puts the frame's floor where the region's hard cases are: Venezuela, Cuba, Haiti and Nicaragua enter with thin statistics, and publishing them as thin evidence is the honest version of the claim.

What the rebase did. At 52 countries, 232 of 276 comparable dimension cells moved against 3.0.0. The mean move is 3.5 points and the largest is 20.3 (Philippines, Agency, 14.2 to 34.5). Brazil moves on six of its seven scored dimensions, Agency most (48.8 to 59.6): the frame now contains the region's weaker states, so middle positions rise. The 3.0.0 numbers are not comparable to these. Every country still scores at least five of nine dimensions, and 0 of 1,604 observed cells clamp.

Cost. Twelve countries with weaker statistical systems thin coverage at the low end: GEM still covers 16 of 52, and the mean move above rests only on the cells scored in both versions. The ingest itself was clean, 7,411 values added with 0 failures and 0 restatements. Sparse-data countries render mostly dashed radars, which is the design working, and A10's warning stands: 52 countries are still not the world.

What would overturn it. Evidence that thin-coverage countries distort an indicator's Tukey fences enough to change well-evidenced countries' readings, or a regional dataset that covers the twelve better than the World Bank does.

D52: A candidate series is tested before it becomes a registry row

Recorded 2026-08-28. Extends D23 and D49.

Choice. pnpm bench probe --series a,b[@db] fetches candidate World Bank series against the live country set and reports four things per candidate: coverage, latest year, spread, and correlation with log GDP per capita. A candidate passes when it covers at least half the country set, carries a value no older than eight years, holds at least three distinct values, and correlates with income below the same 0.70 the diagnostics use. The probe writes nothing. It fetches, prints and exits, so it can run while other work is in flight.

Why. The test already existed and lived in people's heads. A12 records a 25-series probe done by hand against the ten countries of the original prototype, and D44 retired the homicide rate after wiring it, scoring it and reading the diagnostic afterwards. Wiring first and testing later costs a rebase every time it goes wrong, and the country set has since grown from ten to 52, which invalidates the coverage half of every earlier hand probe.

A pass is not a decision. Coverage, recency, spread and the wealth test are necessary and nowhere near sufficient: what the series measures is the argument, and that stays a judgment recorded here. The probe exists to stop that argument being had about a series that has 6 countries or is income wearing a hat.

Cost. One more pipeline module and a command. The thresholds are constants in pipeline/probe.ts and they are arbitrary in the same way every threshold in this project is: defensible, not derived. A candidate that fails coverage today can pass after the publisher's next round, so a failed probe dates.

--search reads the publisher's own catalogue of about 30,000 series by name, so a candidate list starts from what exists rather than from memory. A name absent from the catalogue is absent from the API at every database id, which is how a gap gets recorded as unfillable rather than untried.

Overturned by. A second ingester. The probe speaks World Bank, and the series that would fill Coordination and Trust are the ones the World Bank does not publish, so its long-term job is to be generalised or retired.

D53: The radar answers a pointer, and it has no resting number

Decision. The radar carries its own interaction layer and its own readout. Each of the nine axes owns a sector of the chart as a hit target. Pointing at a sector raises that axis and steps the other eight back, and the number, its confidence and the dimension's question print under the chart.

The readout always shows an axis and never nothing. It opens on the first axis in the fixed dimension order, hover previews another, and leaving the chart returns it to the last axis the reader chose. Every row is always drawn at a fixed height, so the block is the same size before, during and after a hover. Hover and click land on the same axis: a click moves the readout and opens the comparison panel for the dimension the readout is showing.

A finger reads on the first tap and opens the panel on the second. A keyboard reaches the axes as one tab stop and moves between them with the arrow keys. The chart's accessible name is an aria-label, not an SVG <title>, because a <title> renders as a native tooltip over the shape it names.

Why. The chart showed nine numbers and printed none of them. A reader who wanted one had to find the country table further down the page, match a row by name and lose the shape while doing it. Every number was already on the page and none of them was reachable from the picture that plotted it.

The hit target is a sector rather than a vertex because a vertex is two pixels wide. Under a finger it is smaller than the contact patch, and a chart that cannot be read by touch is not readable on the device most Brazilian readers will open it on.

The opening axis is a position in the fixed dimension order and never a summary. The obvious thing to put under a radar at rest is the mean of its axes, and that is the headline score this benchmark refuses to publish. Nine dimensions at equal weights average to a number that ranks countries, which is the claim D16 and D47 exist to withhold. So the readout starts on one dimension out of nine and says which one it is.

Confidence stays a separate statement inside the readout. The score prints through Score, the confidence prints through ConfidenceBar beside it with its band named, and neither is folded into the other.

Cost. Radar is now a client component, so every page that draws one ships it. The grid pages draw 52 of them and pass interactive={false}, which keeps them as pictures inside their links: a hover readout there would compete with the navigation the card exists for, and 52 readouts would add 52 reserved blocks to a page that is meant to be scanned.

The readout is a fixed block under every interactive radar whether or not anybody points at it, and its question line is sized for two lines in both lexicons. A readout that grew and shrank would drag the page under the pointer and make the next axis harder to hit than the last one, so a row that is sometimes half empty is the cheaper of the two failures.

Overturned by. Evidence that readers want the numbers printed on the chart itself. Nine labels on a 260 unit square collide at the sizes this chart is drawn at, which is why they are not there, but a larger chart on the country page could carry them and would make the readout redundant.

D54: An institution map explains capability and never scores it

Recorded 2026-08-28. Extends D20, D31, D35 and D46.

Choice. Each country may publish one institutional capability network at data/institutions/{ISO3}.json. The schema is shared across countries and separates organisations, typed relationships and territorial coverage. A node records the organisation's legal nature, roles, level of government, source and the NCB dimensions it helps a reader investigate. An edge always has a direction, a named relation and its own source. There is no generic connected_to relation.

Official registers provide the structural skeleton where they exist. Brazil uses SIORG for the federal executive and the São Paulo government directory for the state executive. A curated overlay adds the institutions those registers omit and the relationships an organisation chart cannot express: Congress, courts, prosecutors, audit, funding, regulation, data, training and cross-level delivery.

The layer has its own experimental version. It does not change DATASET_VERSION, any DimensionResult, a score or confidence. A direct link from an institution to an NCB dimension is navigation, not a point or a weight. The evidential path stays the one D20 established: an institution participates in a documented delivery, the delivery bears on a declared indicator gap, and that indicator belongs to a dimension.

Brazil begins with a federal baseline, São Paulo as the first state pilot and the municipality of São Paulo as the first municipal connection. Every state and the Federal District is present in the coverage plan from the first file, so the pilot cannot quietly become the scope. Display text is Portuguese for the Brazilian page, while ids, enums and English ground-layer descriptions stay language neutral and translations stay in packages/core/src/i18n/.

Why. The evidence records name actors only inside prose. That is enough to document one delivery and not enough to answer which institutions hold the same capability, which constraints reach them, or what another country would need to reproduce the arrangement. Turning every organisation into an indicator would make the score reward documentation density. Leaving them in paragraphs makes the institutional mechanism impossible to traverse.

A complete force-directed diagram was the obvious first interface and is not the one published. At 47 nodes and 64 relationships it is already a hairball. The viewer opens on one institution, draws its immediate network and lists all of its relationships in words. The directory keeps the whole set reachable, and the data keeps the whole graph available for later views.

Cost. The capability overlay is curated and therefore contestable. SIORG can tell the project that an entity is linked to a ministry; it cannot decide that an audit relation matters to Trust or that a training organisation bears on Learning. Those mappings need sources, review and dates, and the first file is deliberately incomplete outside the federal backbone and São Paulo.

The schema can record that a relation exists but does not yet record its legal scope, intensity or historical intervals. funds can therefore describe a standing funding channel without saying how much moves through it, and two lines with the same verb can carry very different practical weight.

Overturned by. A cross-government official register that publishes the same organisations and capability relationships with stable identifiers, or a cross-country construction whose network measures survive differences in documentation coverage. The first would replace the curated skeleton. The second would justify considering a network measure as evidence, but it would still need a separate decision before entering any score.

D55: Budget execution is the first API-backed Coordination replacement

Recorded 2026-08-28. Extends D45 and D52. Evidence from the World Bank probe and the 4.1.0 ingest.

Choice. Add budget_execution_fidelity to Coordination using World Bank series GF.XPD.BUDG.ZS. The series measures primary government expenditure as a proportion of the original approved budget. The registry applies a new distance_from_100 transform and scores the absolute distance from 100 with lower_better, so both underspending and overspending count as deviation.

The series covers 44 of the 52 countries, reaches 2024, and correlates with log GDP per capita at 0.285 in the pre-wiring probe. After ingest, it gives 44 countries a second observed Coordination indicator beside border compliance time. Those countries now receive a Coordination score under D45. The dataset version moves from 4.0.0 to 4.1.0 because an indicator was added without adding a country.

Why. A country has to carry approved plans into actual spending before it can coordinate public action. This is an observable administrative result, and the API is reproducible. It also improves coverage without restoring the WGI perception composites that made Coordination track income.

Cost. Budget execution is still a proxy. A close match between planned and actual spending does not show that agencies agreed on an objective, delivered it, or produced a useful result. The series has uneven country coverage, its latest value varies by country, and the score remains thin because Coordination still has six other gap or retired rows. The absolute-distance transform is a modelled choice: a small deviation may be healthy under a shock, while a value near 100 may hide poor delivery against a badly designed budget.

Generative estimates remain a separate layer. A gateway or in-session Delphi run may estimate Coordination and Trust where indicator evidence is missing, but generated values stay in delphiScore and blendedScore; they never enter DimensionResult.score, the observation file, or confidence. A real gateway run requires AI_GATEWAY_API_KEY and multi-vendor panel configuration. A mock run exercises the pipeline and is not evidence.

Overturned by. Cross-country delivery evidence showing that budget execution fidelity has little relationship to the ability of independent actors to complete shared objectives, or a better comparable indicator that measures joint delivery directly and survives the same wealth-attribution test.

D56: The institution map reads as a directory and a relation ledger, never as a drawn network

Recorded 2026-08-28. Supersedes the interface D54 described. Extends D26, D35 and D53.

Choice. The institution page publishes no node-link diagram. It publishes three surfaces, each answering one question.

The directory comes first. Filters by name, level and system, then cards grouped by system, each card carrying the institution's level and its number of recorded relations. The reader arrives with a name in mind, so the name is the entry point.

The profile follows, and its relations render as a ledger rather than a picture. Every relation belongs to exactly one of four families declared in INSTITUTION_RELATION_FAMILY in packages/core/src/model/institutions.ts: authority, control, funding and joint work. The families render in a fixed order, and a family with no relation prints that it has none, because an absent relation is a fact about the map. Within a family, incoming relations sit left of a vertical spine and outgoing relations sit right of it, so every line reads left to right in the direction of its own relation and the verb always takes its active form. The geometry supplies the subject.

Nothing on the page is laid out by hand. Every surface derives from counts, so a country with 12 institutions and a country with 400 render through the same code. Institution names live in wrapping DOM text and never inside a fixed-width SVG rectangle.

Language reaches the view as one lex prop, as D53 already requires of the radar. The institution vocabulary moved out of institutions-pt-br.ts and into Lexicon.institutions, so both lexicons carry it and a second country in a second language changes data rather than components. Only the per-institution summaries and the scope sentence stay in the country file, because they describe one country rather than the shared model.

Why. D54 already recorded that a complete diagram of 47 nodes and 64 relationships is a hairball. The published overview grid was that diagram with a grid substituted for a force layout, and it inherited the same failure. Its plane encoded only system membership, which each box label already stated. Its 64 lines rendered 13 different relation verbs identically, so the picture asserted only that the institutions are connected, which curation guarantees. Two nodes held 33 of the 128 relation endpoints, so the layout spent its whole canvas drawing the spokes of a star. Node labels ran at 9px and truncated at 16 characters, below anything in the type scale, and the mobile branch rendered no diagram at all.

The families are the move that makes the ledger carry information a list of lines could not. They match the three questions the page headline already asks, and they let an empty band speak: BNDES exercises one relation and receives three, and the ledger shows that shape at a glance.

The same day this was recorded, the Brazilian file grew from 47 institutions toward a curated entry for every state. A layout with hard-coded columns and fixed box heights would have broken on that growth. A count-derived layout did not.

Cost. The page no longer offers any single picture of a whole country's wiring. A reader who wants to see the shape of the state as one object has to assemble it from profiles. The system-by-system relation matrix is the intended answer and is not built: in the current Brazilian file 30 of its 100 cells are filled, which is dense enough to read and sparse enough to be legible, but it is a separate change.

The level enum still carries Brazilian assumptions in what its four values mean. Germany, the United Kingdom and a European Union member state each divide government differently, and external describes a different relationship in each. The enum survives the first international map only because the labels are now a lexicon lookup.

Overturned by. A country map whose relation set does not sort cleanly into the four families, which would mean the families encode Brazil rather than the model. Or evidence from readers that the whole-network shape is the question they arrive with, which would make the matrix the page's first surface rather than its missing one.

D57: Trust is two families, and it publishes nothing until both are answered

Recorded 2026-08-28. Extends D20, D23, D44 and D45. Adds the indicator family to the registry.

Choice. Trust means the expectation that people and institutions behave reliably outside close personal networks. That is two questions, not one, and the registry now says so. Every Trust row carries a family: social for generalised interpersonal trust and trust in strangers, institutional for confidence in government, courts and the civil service, contract enforcement time and court case clearance. court_case_clearance is added as a gap in the institutional family, because D23 named court throughput as the observable replacement for the two retired WGI composites and nothing has filled it since. homicide_rate stays retired and stays untagged: D44 retired it because it belongs to neither family.

The family weights nothing. Scoring stays the equal-weight mean of whatever is observed, and the coverage floor from D45 still decides whether a dimension publishes. The family is a diagnostic. familyBalance in packages/core/src/pipeline/diagnostics.ts reports, per dimension that declares families, how many indicators each family holds, how many are observed, and how many scored countries rest on a single family. The report and the diagnostics page render it.

A first credible Trust score is one harmonised social measure plus one comparable institutional-performance measure. Until both exist across enough countries, Trust publishes no score.

What this rules out. The WGI rule of law and control of corruption composites stay retired: they correlate with each other above 0.95 and with log GDP per capita at 0.83, and they are one measurement wearing two names. Homicide stays retired: physical safety is not trust, and D42 measured it adding 0.288 to this dimension's wealth correlation. IC.FRM.CORR.ZS is ineligible on its definition, which asks a firm what it believes firms similar to itself pay, so it records belief rather than experience. No synthetic or model-generated value ever becomes an observation.

IC.FRM.BRIB.ZS is the one Enterprise Survey series this decision does not reject on its definition. It asks whether the responding firm was itself asked for a bribe across six public transactions, so it records experience. It covers 49 of 52 countries, 44 of them at 2023 or later. It is admissible as a secondary behavioural check and it is not admissible as the Trust score, because a rank-normalised estimate puts it at about 0.66 against log GDP per capita on its own and puts the two-indicator dimension at about 0.53, against 0.14 for contract enforcement days alone. That is a larger wealth contribution than the one D44 retired an indicator over. Scoring Trust with it would clear the coverage floor by re-creating A3 and A4.

Why. Trust was the one dimension where the fastest route to a published number was also the route that would make the number wrong, and the pressure to take it came from the empty axis rather than from any evidence. Writing the two families into the registry makes the empty axis legible: it is not that Trust has one indicator, it is that Trust has one family answered and one family with no data at all. It also states the acceptance test in advance, so a future series is judged against a written contract instead of against the wish for a complete radar.

The families exist as a diagnostic rather than as a weight because weighting them now would be a second unevidenced choice on top of a thin dimension. Three correlated survey items counted as three independent signals is a real problem, and it is a problem that starts when the data arrives. The count is published first so the weighting decision, when it comes, can cite it.

Costs. One more gap row lowers Trust's confidence again, which is correct and reads as a regression to anybody who has not read this entry. The family is a free-text field on the registry, so a typo makes a new family silently. Nothing validates the vocabulary, because a fixed enum would have to guess the families of eight other dimensions that have not needed them yet.

Overturned by. A harmonised social-trust series and a comparable institutional-performance series that together clear the floor, whose combined dimension survives the D42 wealth-attribution test. That closes the decision by satisfying it. The decision is wrong instead if the two families turn out to correlate above the redundancy threshold once both are measured, which would mean Trust asks one question after all and the split encodes a distinction the world does not make.

D58: The country's wiring is a system matrix, and its ramp is fixed rather than fitted

Recorded 2026-08-28. Completes D56, which named this surface and did not build it.

Choice. The institution page publishes one picture of a whole country: a matrix counting the relations that run from every system to every system, ten by ten, including the diagonal. buildInstitutionMatrix in packages/core/src/pipeline/institutions.ts computes it, and INSTITUTION_SYSTEMS in the model fixes the axis order so two lexicons cannot present the same matrix with its rows in different places. Nothing here reaches a score, a confidence or data/out.

A family filter narrows the count to authority, control, funding or joint work, which is where the matrix earns its place: the four families run through visibly different cells, and the whole-map view hides that.

The matrix owns its readout, the way the radar does. It always reads one cell, it opens on the busiest cell rather than on nothing, and the readout names both institutions in every relation as links into their profiles, so the matrix is a way into the map and not only a summary of it.

The ramp is three fixed breaks on the count: one, two to four, five or more. Colour never encodes the number, which is printed in the cell; it encodes whether a channel is busy. The ramp is the score ramp with its lime top removed, because a hundred lime cells would spend the single accent this page has.

Why. Terciles were tried first and are wrong for this quantity. Relation counts are one spike at 1 with a thin long tail: in the current Brazilian map 13 of 33 filled cells hold exactly one relation and the busiest holds 52. Cutting at terciles left the bottom band empty, put 2 and 52 in the same band, and gave the joint work view a single band covering 1 to 26. A fitted ramp also moved whenever the family filter changed, so the same colour meant a different quantity one click later.

Fixed breaks are available here in a way they are not for a score. A count is absolute. One relation is one relation in every country, while a score is a position inside a frame that a new country rebases. So the ramp holds still across families and across countries, and two national matrices can be read side by side. Against the current file every family view fills all three bands.

The matrix renders as a <table> with row and column headers rather than through DataTable. DataTable owns its <td> and sorts columns; a matrix sorts nothing and its cells are buttons. The table element is what carries the row and column relationship to a screen reader, and that is the reason to use it.

Selecting an institution now scrolls the profile into view. The Brazilian directory passed two hundred cards while this was being built, far enough that a click near the bottom changed a heading the reader could not see.

Cost. A cell counts relations of very different weight, which is the limit D54 already recorded in the schema: funds can describe a standing channel without saying how much moves through it. A busy cell therefore means well documented as much as it means important, and the matrix inherits every curation bias in the map beneath it.

Aggregating to ten systems also hides which institution inside a system holds a channel. Two hundred institutions collapse into a hundred cells, and the readout is the only way back out.

Overturned by. A country whose map fills so few cells that the matrix reads as empty rather than sparse, which would mean the surface needs a coverage threshold before it renders. Or relation weights in the schema, which would make a count the wrong quantity to put in a cell.

D59: The data publishes a quoting contract for automated readers, and no MCP server

Choice. Publish two agent-facing surfaces and stop there. docs/FOR-AGENTS.md states the contract: a score never travels alone, which fields must accompany it, and the six things not to do with it. /llms.txt in the viewer is the map that points at it, generated from the registry and the current index rather than hand-written. Do not build an MCP server for this dataset yet.

Why. The risk this project runs with an automated reader is not that the data is hard to reach. It is that the data is easy to reach and the caveats are not attached to it. index.json hands a model 52 countries and nine scores in one request. Nothing in that file stops the model averaging the nine, ranking the countries by the mean, quoting a very thin dimension on its own, or reading a null as a zero. Every one of those is a claim the project explicitly refuses to make, and every one of them is one fetch away.

So the gap is a contract, not a transport. A written contract closes it at the cost of one file, and the same file serves a human reading the repository.

/llms.txt is generated because the alternative is a hand-kept file stating a country count and a dataset version that the next re-ingest moves. A stale map of a dataset that versions itself is worse than no map.

An MCP server was the other candidate and it is the wrong tool here now. The whole scored output is a few megabytes and index.json is about half a megabyte, so there is no corpus too large to hand over and no query problem to solve. The directory is already self-describing through datapackage.json and the generated schemas, so a client that can fetch a URL can already read it and validate what it read. Nothing is behind auth. Against that, a server is a second place the published shape is written, which is exactly what D30 and the single-source-of-truth invariants exist to prevent, and it is a deployment that can drift from the dataset version it wraps.

Cost. A contract in a file is advisory. A server could refuse to answer without the confidence attached; a markdown document can only ask. We are choosing the surface that cannot enforce, on the argument that a reader who will ignore docs/FOR-AGENTS.md will also strip the fields an MCP tool returns.

/llms.txt is also a convention rather than a standard, and it is read by some clients and ignored by most. The file is cheap enough that this is acceptable, but it should not be treated as coverage.

Generating the file makes it a dynamic route rather than a static asset, so it costs a function invocation and it fails if data/out is missing, the same way every other page here does.

Overturned by. Evidence that agents are a real consumer rather than an assumed one: referrer or user-agent data showing automated fetches of data/out, or a citation of an NCB number in a generated document that drops the confidence. Either would justify the enforcing surface. The other trigger is the Delphi layer becoming interactive, so a caller wants to run a scored what-if against the frame instead of reading a fixed file. That is a tool call, not a document, and it is the point where MCP earns its keep. It belongs as a route handler in apps/web reading @ncb/core and data/out the way the pages do, so that no third copy of the model exists.

D60: A series that cannot be scored is published beside the score, with the reason attached

Recorded 2026-08-28. Extends D23, D42, D44 and D57. Adds the behavioural check to the model.

Choice. The model gains a third kind of row. An indicator is scored. A gap is declared and unfilled. A check is fetched, published and excluded from every number: the frame, the mean, the coverage floor, the indicator count and the confidence. Checks live in packages/core/src/model/checks.ts, they are observed under the __check__ prefix in the observation file, and each one carries the reason it is not scored as a field that renders to the reader.

The first check is bribery_incidence, World Bank IC.FRM.BRIB.ZS, filed against Trust in the institutional family. It covers 49 of 52 countries and 44 of them at 2023 or later. It asks whether the responding firm was itself asked for a bribe across six public transactions, so it records experience rather than reputation, which is what separates it from the perception composites D23 retired and from IC.FRM.CORR.ZS, whose own definition asks a firm what it believes firms similar to itself pay.

It is not scored. On a rank-normalised estimate it correlates with log GDP per capita at about 0.66 by itself and takes the two-indicator Trust dimension to about 0.53, where contract enforcement days alone sits at 0.14. That is a larger wealth contribution than the 0.288 D44 retired an indicator over, so scoring it would clear the D45 coverage floor by re-creating A3 and A4. D57 already fixed what Trust needs before it publishes, and this series is not it.

behaviouralChecks in the diagnostics recomputes the wealth test on every run, so the exclusion is standing evidence rather than a claim in a document. A check is published as the publisher wrote it and is never normalised, because putting it on the 0 to 100 scale would invite exactly the reading this decision refuses.

Why. Retiring a series and publishing a series were the only two options the model had, and neither fits a number that is real, current, wide in coverage and disqualified. Retiring it hides evidence a reader should see. Scoring it makes the benchmark assert something its own diagnostics reject. The check is the third option: show the number, show the reason, keep it out of the arithmetic.

It also changes what an empty dimension looks like. Trust publishes no score and now publishes a current, behavioural, 49-country reading beside the empty axis. That is a more honest surface than either a blank or a number the model does not believe.

Costs. A published number that is not in the score will be quoted as though it were, and no design prevents that. The mitigation is that the reason travels in the same object: note on every published CheckResult, rendered on the country page and the capability page and printed in the report. A second cost is that the check is a new concept with a small surface, which makes it a place to put anything inconvenient. The rule against that is in checks.ts: a series that passes the tests belongs in indicators.ts, and a check needs a decision entry naming the test it failed. Third, the /sources page is built from the indicator registry and does not yet list check series, so the fetch it prints back is now incomplete by one call.

Overturned by. Evidence that readers treat a check as a score, which would mean publishing it beside the dimension does the harm the exclusion was meant to avoid, and the row should be retired instead. Or a change in the series that breaks its correlation with income, which would make it an indicator and move it to the other registry.

D61: Social cards are static metadata images with stable public paths

Recorded 2026-08-29. Implements issue 1.

Choice. Generate the country, capability and agenda social cards through Next's opengraph-image metadata convention and expose them at the stable paths /og/country/<ISO3>, /og/dimension/<dimension> and /og/agenda/<ISO3> with rewrites. Each image uses ImageResponse, reads the same scored output as the viewer, and declares generateStaticParams, so the cards are emitted during the production build and served as immutable assets.

Radar coordinates live in a browser safe module shared by Radar and the server-only radarToSvgPath entry point in Og.tsx. Confidence is passed to the helper for profile parity but does not change a score's position. The dimension card pins Brazil as the reference country because that route has no country parameter of its own.

The current dataset emits 52 country cards, 52 agenda cards and nine dimension cards, 113 images in total. The issue description's arithmetic of 52 cards times three routes would imply 156 images, but the dimension route has nine valid parameters, one per capability dimension.

Why. Social crawlers need a stable image URL and the benchmark's radar is the clearest compact representation of a country profile. Build time generation keeps the card independent of request latency, gives the CDN an immutable asset, and makes a missing country or dimension fail during the build rather than when a reader shares a link.

The image renderer cannot parse the page's WOFF2 fonts, and its default emoji renderer fetches flag artwork from a remote CDN. The cards therefore register Next's bundled TTF and use the registry's ISO2 code in a lime country mark. This keeps builds offline and reproducible. Vendored flag artwork can replace the mark later if exact national flag rendering becomes material.

Costs. The image route is a second presentation of the data, so its visual layout can drift from the viewer even though its radar coordinates are shared. The card also has less room for confidence and provenance than the page, which is why it carries the site level caveat and never introduces a headline score. Static output must be rebuilt after a dataset refresh, just like the committed JSON and agenda documents.

Overturned by. Evidence that crawlers cannot follow the stable rewrites, or that shared cards are stale after a data release, would justify an on demand route or a different cache strategy. Evidence that readers need exact national flag artwork would justify adding a small vendored asset set rather than reintroducing a remote emoji dependency.

D62: Disputes preserve the target snapshot and count distinct target countries

Recorded 2026-08-29. Implements issue 2.

Choice. A score can open a small challenge form with its country, dimension, score and confidence already attached. The POST endpoint validates the argument, checks the country and dimension against the registry, then reads the current country file to store the canonical target values. The record is appended to data/disputes/<YYYY-MM-DD>.jsonl with status: submitted. The schema also holds a maintainer response and signature, and requires a signature before an accepted record can parse. A rejected record remains in the public ledger but does not count toward a contested badge.

The badge threshold is three non-rejected disputes from three distinct target countries in one dimension. It appears only on the target cells named by those disputes, and carries the distinct-country count. The threshold is a constant in packages/core/src/model/challenges.ts rather than a display decision hidden in a component.

Why. A reader should be able to challenge the number while looking at it, and the challenge should keep the exact score and confidence that prompted the argument. Counting distinct countries prevents repeated submissions from one place from looking like independent corroboration. Keeping rejected records visible preserves the review trail without letting a decision that was declined change the page's warning state.

Costs. The current ledger is an append-only filesystem store. It works in a checkout and on a self-hosted Node process, but a Vercel deployment needs a durable write target before this endpoint can accept public submissions at scale. Acceptance remains a maintainer action: changing status, adding a signature and writing the resolving decision are intentionally separate from a reader's POST. The form also accepts an optional URL, so an argument can still be weak even when it looks complete.

Overturned by. Evidence that the filesystem ledger loses accepted records or receives enough submissions to need concurrency control would justify a durable store and an authenticated review tool. Evidence that three distinct countries does not separate useful disputes from repetition would justify a new threshold decision with observed submission data.

D63: Source-backed scores and Delphi estimates remain separate tracks

Recorded 2026-08-29. Clarifies D11 and supersedes the wider fallback described as a cost in D45.

Choice. The source-backed track is the measurement layer. It uses the registry, named publishers and observed values to produce score, confidence and indicator rows. World Bank is the only automated ingestion source in v0; manually authored observations remain a separate input. The Delphi track is an interpretation layer. It reads the source-backed evidence brief to review thin or questionable dimensions, then stores delphiScore, delphiIqr, rationales, self-confidence and missing evidence. It never creates an observation, changes confidence or enters DimensionResult.score.

blendedScore uses the indicator score when the dimension clears its coverage floor. It falls back to Delphi only when no indicator is observed, and blendedFrom records the source. A dimension with one observed indicator stays unmeasured rather than receiving a generated replacement.

Runs record the dataset version, country set, scope, coverage ceiling and prompt version alongside model identity and provenance. A country-restricted or coverage-restricted run is an archive by default. It becomes active only when the operator passes --activate. A published run after a country-set change must cover the full rebased set.

Why. The benchmark needs a number that can be traced to a source and a separate judgment about what that number misses. Joining them would hide the difference between measurement and interpretation, and would make a model estimate look like a new observation. Explicit run scope also keeps a 10-country preflight from replacing the active run for the whole benchmark.

Costs. Some dimensions remain without a blended value when one indicator is present but the coverage floor is not met. A full panel rerun is required after a country addition if its estimates are interpreted against the rebased frame. The panel remains an audit and fallback, not a substitute for closing source gaps.

Overturned by. A defensible method that combines source and panel values without hiding either provenance, or a source-backed series that measures the currently unmeasured dimensions with comparable country coverage, would justify a new decision.

D64: The first Trust social measure is a pinned Joint EVS/WVS adapter

Recorded 2026-08-29. Extends D57 and D63.

Choice. The first source-backed social measure for Trust is A165, “Most people can be trusted”, from the official Joint EVS/WVS 2017-2022 results release 5.0.0. The adapter reads the publisher's aggregate table, which is weighted by gwght, preserves the published percentage answering 1 (trusted), stores the release year 2022 and emits the existing observation shape. It does not copy respondent-level microdata or ask a model to generate a value.

The adapter returns the shared coverage report defined in pipeline/adapters/types.ts, writes data/observations/joint-evs-wvs.json and records additions or changes in data/observations/revisions.json. The store loads it alongside World Bank and manual observations, so scoring, confidence, diagnostics and reports use the same metric path for every source. Countries with separate EVS and WVS rows, currently Germany, Great Britain and the Netherlands, are held until pooled microdata weights can be harmonised reproducibly. The existing contract-enforcement row is the institutional- performance leg for this provisional release; court clearance remains a gap.

Why. The official aggregate release is inspectable, has a stable source endpoint and covers the benchmark sufficiently for the first social-family release. Using the publisher's weighted result avoids silently choosing a pooling rule for two survey programmes and avoids storing restricted respondent-level data. A source adapter makes the extraction repeatable and lets future EVS/WVS releases replace the file through the same revision log.

Costs. Only 36 of 52 benchmark countries emit a unique row, although 39 are recognized before duplicate-country rows are held. The item is a perception proxy, its reference year represents a multi-year fieldwork release, and the 2019 contract-enforcement series is stale. Trust therefore publishes a thin two-indicator score for 36 countries at confidence 0.159. Its current GDP correlation is 0.627, and the A165 row alone is 0.669, so the release remains a watch item for wealth sensitivity, survey comparability and redundancy.

Overturned by. A reproducible pooled-microdata treatment that improves coverage without weakening comparability, or a comparable court-performance series that changes the institutional leg, would justify a follow-up decision. Evidence that A165 is not comparable across the benchmark countries or that the two-family Trust result fails the wealth and redundancy review would require retiring or reclassifying the row.

D65: Provisional layers need evidence before public promotion

Recorded 2026-08-29. Implements issue 14 and sets the promotion gate for issues 11, 12 and 13.

Choice. Velocity and Exponential Leverage remain provisional and offline until each layer meets its own promotion criteria. The criteria are:

1. Velocity has a settled method. The base and current years, the treatment of negative deltas, low-confidence dimensions and confidence changes are documented, tested against the fixture, and stable across a six-month record of three quarterly reviews under the process in issue 13. At least one external reviewer must confirm that the result is interpretable as a rate of movement rather than a second capability score. 2. Exponential Leverage has a settled method and complete shape. All eleven dimensions are either sourced or explicitly labeled metric-under- development, and the weights, offsets, source changes and foundation coherence rule are documented and tested. The layer must also have a six-month record of three quarterly reviews and at least one public review from outside the maintainer team. 3. Promotion is a recorded decision. Passing the technical checks does not publish either layer. A later decision must link the review record, the external or public review signal, the final fixture and the user-facing caveat before the sandbox becomes a public surface.

These criteria are the contract between the provisional fixtures in issues 11 and 12 and the quarterly review process in issue 13. Until the criteria are met, a missing value is a missing value, and neither layer is used in the headline score, confidence, agenda, or ranking.

Why. Both layers are useful experiments but can look more authoritative than their inputs deserve. A fixed review period makes learning visible, while the separate requirements force the two layers to resolve their different methodological risks. External review is necessary because internal iteration can show that code runs without showing that the measure means what readers think it means.

Costs. The layers may remain provisional for longer than a release cycle, and a useful sandbox result may never qualify for publication. The review record also creates maintenance work every quarter. Those costs are preferable to turning an exploratory rate or composite into an unqualified benchmark claim.

Overturned by. Evidence that the criteria do not detect methodological failure, that three quarterly reviews are too short to reveal instability, or that independent reviewers cannot distinguish either layer from a capability score would require revising the promotion gate before publication. Evidence that a layer is not learning after three consecutive reviews would trigger retirement rather than a weaker promotion standard.

D66: Subnational observations corroborate the national comparison layer

Recorded 2026-08-29. Extends D12, D25 and D49. Implements issue 9.

Choice. The benchmark keeps one comparison layer: observations with geometry: "national" are the only observations read by frames, scores, trends and confidence. The observation schema also admits state, province, region and municipality rows, each carrying a reconciliation rule of aggregate, independent or context_only. Existing files default to the national geometry and the context-only rule when they do not yet carry the new fields.

Subnational rows live beside the comparison layer and are consumed by a destination page through a separately validated corroboration fixture. The fixture carries the publisher's national value, constituent values, source, date and rule together. It does not produce a per-state capability score or silently reweight the national result.

Why. A federal aggregate answers the cross-country question but can hide variation between the units that actually deliver policy. Adding a geometry field makes that missing layer explicit without allowing a state observation to enter a national frame. Requiring a reconciliation rule prevents a reader from assuming that every state range can be averaged into a national statistic.

Costs. The data model has two spatial layers to maintain, and every subnational fixture needs its own source, date and editorial review. A fixture can show that a national number is consistent with its parts, but it cannot make unlike administrative units comparable or turn a context measure into a score. The initial reader and fixture are intentionally limited to Brazil.

Overturned by. A documented method that establishes a comparable subnational capability frame without changing the national comparison, or evidence that the declared reconciliation rules systematically mislead readers, would justify revising this boundary. A source that publishes only incompatible geometries would justify withholding its fixture rather than relaxing the schema.

D67: The front page reads one capability at a time, and certainty is the ring

Recorded 2026-08-30. Extends D32, D47 and D53.

Choice. The front page opens on a single capability, drawn as every country on one 0 to 100 axis, and the reader switches capability rather than scrolling past nine charts. One component, FlagHistogram, serves all nine dimensions. Switching moves each country from its old score to its new one instead of redrawing the field. The frame is laid out for all nine dimensions at once, so the chart height and the axis hold still while the flags travel.

Pointing at a flag opens a card beside it that reads the country at a glance: the score on this capability, its confidence, its trend, and the highest and lowest of the same country's nine capabilities. The card never takes the pointer. The flag itself is a link into the country profile, so the card can stay purely informative and a reader never has to travel into a tooltip to click something. No flag fades while another is read: the halo says which one is being pointed at, and 52 flags at two strengths reads as a rendering fault.

The chart itself is FlagField, and it is the only one of its kind. The capability pages, the dimension and indicator peek panels and the compare embed all draw the same picture with the same geometry, from the same component. A surface that needs more in the hover card adds a field to FlagFieldPoint rather than a second distribution. Distribution, the older strip chart that mapped confidence to marker opacity, is deleted.

The mark is FlagBubble: a flag inside a bubble, drawn at the origin so the chart owns the position. The bubble's ring carries the evidence and the flag never does. The ring is solid at the usable band's floor and above, and it breaks below it, with the gaps opening as confidence falls. The ramp lives in apps/web/src/lib/evidence.ts as evidenceOpenness, and the radar's dashed edge reads the same function, so the two charts cannot disagree about which countries are drawn solid.

Each dimension names both ends of its own axis. DIMENSION_ENDPOINTS in packages/core/src/model/dimensions.ts is the only place those words are written, beside the questions they belong to.

The grid of 52 radars moves to /countries, which the Countries nav entry now points at, and the full score table is added to /capabilities. The front page keeps the dataset markup and gains no ranking.

Why. The old front page asked a reader to compare 52 nine-axis shapes before knowing what any axis meant. A shape is the right picture for one country and the wrong one for a first visit. One capability on one axis is readable without instruction, and moving the flags between capabilities is the only way the page demonstrates its own claim: that a country strong on one capability can sit at the floor of the next. That claim is the reason there is no composite score.

Opacity was the wrong channel for certainty. It asks a reader to compare the strength of 52 marks against different backgrounds, it collides with the dimming used for hover, and on a flag it reads as a rendering fault rather than as a measure. A ring is a separate channel from the mark it surrounds, and it already means thin evidence on the radar.

Costs. The front page now shows one dimension rather than nine, so a reader who wants the whole picture makes one more click. The animation needs every country's nine scores in the client bundle, which is the slim index and not the country files. The shared field is taller than the strip chart it replaces, because a bubble needs more column spacing than a bare flag, so the compare embed's advertised height rises from 300 to 420 and anybody who has already embedded it keeps the old height until they update the snippet.

Overturned by. Evidence that readers cannot find the capability switch, or that they read the animated move as data changing rather than as the question changing, would justify returning to nine small charts shown together. Evidence that a broken ring is not read as uncertainty in testing would justify a different second channel, but not a return to flag opacity.

D68: The wealth residual is published per dimension and never summed

Recorded 2026-08-30. Extends D1, D23 and D47. Joins the provisional gate in D65.

Choice. A new provisional layer, written by pnpm bench residual into data/out/residual.json under residual/0.1-exploratory. For each dimension, ordinary least squares fits the published score against log10 GDP per capita across every country scored on that dimension. The residual is the observed score minus the fitted score. The layer publishes:

• one residual per country per dimension, beside the score and the fitted value,
• one fit per dimension carrying slope, intercept, Pearson r, r², n, the standard error of the estimate, a fit strength band and the mean absolute rank shift between the score order and the residual order.

The shape has no country-level field and never will. Nine residuals averaged into one number is the headline score D1 withholds with a regression in front of it. Residuals never enter score, blendedScore, confidence, momentum or the agenda. A dimension with no score publishes no residual, a country with no income observation publishes none at all, and a dimension fitted on fewer than 20 countries publishes neither a fit nor its residuals.

The layer is offline in the sense D65 means. It renders at /method/residual/sandbox, which is unlinked and carries robots: noindex, the same shape as the velocity and leverage sandboxes. No score, confidence, agenda, country page or public surface reads the file, and no navigation points at the sandbox until a later decision promotes it.

Why. D1 refuses a composite because the mean of nine dimensions ranks countries. Measured on the current release, that mean correlates with log GDP per capita at r 0.866, Spearman 0.879, higher than seven of the nine dimensions on their own. Averaging cancels the two dimensions that do not track income, Shared Purpose at 0.451 and Coordination at 0.467, against the ones that do, and what survives the cancellation is wealth. The mean is also computed over different baskets: 23 of 52 countries score fewer than nine dimensions under the D45 coverage floor, so a mean over five and a mean over nine would print as the same kind of number.

The residual answers the question the composite was reaching for and does not reproduce the development ranking. It says whether a country is above or below what its income buys, which is the comparison this benchmark exists to make. The first run puts China +21.6, South Korea +11.4, Finland +10.2, Rwanda +9.6 and Ethiopia +8.9 above their income lines on the mean of the nine residuals, and Panama -17.2, the United Arab Emirates -16.8 and Argentina -11.1 below. Those five numbers are stated here as the evidence for the layer and are exactly what the published file refuses to compute, because the same averaging failure applies to residuals.

The fit travels with every residual for one reason. Where the income line explains little, the residual is close to the score, and a weak fit dressed as a residual would smuggle a capability ranking back in. Coordination and Shared Purpose are both weak on this release, at r² 0.218 and 0.204, and their mean absolute rank shifts are 6.9 places out of 44 and 5.5 out of 46. Anticipation, at r² 0.763, moves 11.9 places out of 50. The layer publishes the difference rather than hiding it.

Costs.

• A regression sits between the reader and the data. Every published residual depends on a modeling choice, which no score in this benchmark does.
• The country being read helped fit the line it is measured against, so an outlier pulls the line toward itself and understates its own residual. China is the case to watch. A leave-one-out fit is the candidate fix for 0.2.
• The residual inherits A3. Where a dimension's evidence is itself wealth correlated, removing income removes part of the capability signal too.
• A negative residual reads as blame. "Below what its income predicts" is a statement about a fit, not about effort, and the layer has no user-facing copy yet that says so.
• The mean absolute rank shift is in places, so it is read against n in the same row and is not comparable across dimensions with different country counts.
• A straight line in log income predicts scores the scale cannot hold. Five of 424 cells fit below zero, all of them Ethiopia, Haiti and Rwanda on Experimentation and Anticipation, and the residual there carries the impossible part with it. The cell records this as outOfScale, the same convention a clamped scored cell uses with outOfFrame. Clamping the fitted value would break the property that makes the residual worth having, because residuals of a clamped line no longer sum to zero against income.
• Venezuela and Cuba have no GDP per capita observation in the current ingestion, so they carry no residual on any dimension.

Overturned by. Evidence that residual order tracks income anyway. Pearson against log GDP per capita is zero by construction under a full-set linear fit, so the test is the rank correlation, which is not forced. On this release every dimension sits between -0.081 and +0.129 in Spearman terms. A later release carrying a dimension above about 0.3 would mean the relation between income and that dimension is not linear in log income, and a straight line is the wrong model for it. Evidence that readers read a residual as a capability score would stop promotion under D65 rather than change the method. Evidence that leave-one-out fits move outlier residuals by more than the standard error of the estimate would make the full-set fit in 0.1 wrong rather than simple.

D69: A country layer replaces the translated edition

Recorded 2026-08-30. Supersedes the viewer half of D35.

Choice. The viewer publishes one benchmark, in English, and country layers beside it. A country layer is a second reading of one country, written in that country's language and at that country's depth: its shape, its computed agenda, its institution map and its subnational spread. It is not the benchmark translated. Brazil is the first and today the only layer, at /brasil.

COUNTRY_LAYERS in apps/web/src/lib/layers.ts is the only place a layer is declared: the country, the slug, the language, the section order and the labels. Every address, the nav between the layer's pages, the way back out to the comparison, and the gate on which languages a country's pages may render in, all read that registry.

The Portuguese mirror is gone: /pt, the Portuguese agenda list over all 52 countries, and the translated method, glossary, limits and decisions pages. Every one of those addresses now redirects into the layer or back into the ground layer. The home page no longer changes language with Accept-Language, the country agenda no longer honours ?lang=, and the header carries no language control. A lexicon still renders any country, but a rendered lexicon with no page behind it is not a published thing, so the feed drops those entries.

No country layer appears in the sections at the top of the site, and no layer has a nav of its own. A layer serves one country's audience and is reached from that country. Where it sits in the navigation is settled by D73, which gives the whole viewer one tree: a layer is a reading of its country, beside the English one.

The ground-layer half of D35 stands unchanged. Ids, registry definitions, JSON output and the method documents stay English, and the layer reads exactly those files, so a claim it makes can be checked against its source.

Why. A translated edition promises a second complete site and cannot keep it. Ours mirrored nine pages of a growing site, and the mirror drifted the moment either side moved, while the pages a Brazilian institution actually needs, the institution map and the state spread, had no home in either language. A layer inverts that cost: it grows only where the project has done country-specific work, and its existence is evidence that the work was done. It also states the audience honestly. The benchmark's subject is the comparison, and a reader who wants Brazil is a sub-audience, reached from Brazil rather than from the top of the site.

Cost. Brazil's pages exist twice, once in the comparison and once in the layer, and the two can disagree in emphasis even though they read the same JSON. A Portuguese reader on any other country now gets English with no offer of anything else, which is a real loss for the 51 countries that will not get a layer. pnpm bench agenda still renders {ISO3}.pt-BR.md for every country, so those files are generated with no page behind them; restricting generation to layer countries is separate work. The layer's subnational section is still the English ground-layer page at /country/BRA/local, declared in the registry with a null slug until it is written in Portuguese.

Overturned by. A second lusophone or hispanophone country layer whose sections turn out to be identical to Brazil's, which would show that what we called country-specific work was a language after all. Evidence that readers arriving from a Portuguese search reach the English home and leave, which would justify an entry offer on the ground layer rather than only inside Brazil's pages. A partner institution that needs the method and limits documents in Portuguese in order to review the framework, which would justify translating those two documents as documents rather than restoring an edition.

D70: A comparison is an address, and one country in it is the reference

Recorded 2026-08-30.

Choice. The viewer publishes a top-level comparison at /compare. It holds one reference country plus up to three others, and the whole selection lives in the path: /compare/BRA-IDN-ZAF. The first code is the reference. Every other column is read as a distance from it, and moving a country to the front changes the reading without changing a number.

compareHref and readCompareCodes in apps/web/src/lib/links.ts are the only places the shape is written or parsed, as D46 requires. The parser is forgiving in the ways a person writing a URL is: hyphens, slashes and commas all separate, case does not matter, and a code that is not in the country registry is dropped rather than raised. The page then redirects to the canonical hyphen form, so a hand-typed address and a shared one are the same page. COMPARE_MAX is four.

The picker writes the address and never local state, so every selection a reader builds is a link they can send. The page reads the country files rather than the index, because it goes down to the indicator rows; four files is a bounded read and never the list D27 forbids.

Four shapes are not stacked on one radar. Each country gets its own card with the reference drawn behind it as a muted outline. Comparison stays in the tables: nine capability rows with the gap from the reference under each score, confidence in its own table, trend in a third, and then every indicator in the registry with each country's normalised chip and published value.

Reference country is a glossary entry, because the page invents the term.

Why. A score in this benchmark only means something against other countries, and until now the viewer offered two ways to get there: one comparator behind a selector on a country profile, which vanished when the reader navigated away, and one capability across all 52 countries. Neither answers the question a reader actually arrives with, which is how a small set of countries differ across the whole profile. Putting the selection in the path makes that reading a citable object: a researcher can send /compare/BRA-IDN into a document and it opens the same page for everybody.

The reference is named rather than implied because a table of four columns with no stated baseline invites the reader to invent one, usually the highest column, which is the ranking this project withholds. Naming it makes the arithmetic explicit and reversible.

Cost. The page is the only surface that loads several country files at once, so it is the heaviest page in the viewer, and the indicator section grows with the registry. A four-country comparison is wide on a phone: the tables scroll horizontally inside their own containers rather than reflowing. The reference also carries a risk the picker cannot remove, which is that a reader treats the reference country as a target instead of a baseline. The glossary entry and the column label say it is not; nothing enforces it.

Combinations are not in the sitemap. 52 countries make 1,326 pairs before any triple, and a crawl map that carried them would drown the pages that are actually published. Only /compare is listed.

Overturned by. Evidence that readers build comparisons and never send them, which would make the address a cost with no benefit and argue for local state and a shorter page. A reader study showing that the reference column is read as a target would force the gap column out, leaving four independent columns. A fifth country asked for often enough would mean the page is being used as a table, and the answer to that is the flat CSV rather than a wider page.

D71: One inbox, and support is a page on both layers

Recorded 2026-08-30.

Choice. The viewer publishes one communication page, /contact, and one support page per layer: /support on the ground layer and /brasil/apoie inside Brazil's layer. supportHref and contactHref in apps/web/src/lib/links.ts are the only places those addresses are written, as D46 requires, and support joins LayerSectionId in apps/web/src/lib/layers.ts, so a layer either holds its own reading of the page or falls back to the ground-layer one through the same map every other section uses.

The support pages name three things and no more: use the benchmark, contribute to it, fund a named piece of it. Each way ends at an address a reader can act on today, and every invitation to write ends at /contact carrying a topic in the query string.

/api/contact validates ContactSubmission from packages/core/src/model/contact.ts, redeems a Turnstile token when the keys are set, and forwards the lead to core.envisioning.com over an HMAC-signed request through apps/web/src/lib/core-api.ts. It stores nothing in this repository. The payload is the one envisioning.com and event-bff already send, including the two traps those repositories document: sourcePage is never sent, because Core answers 500 when it is present, and an empty title is replaced with a placeholder, because Core rejects a blank one. Without INTERNAL_REQUEST_SECRET the route answers 503 rather than pretending to send.

Why. A benchmark that asks institutions to use it, argue with it and fund it has to say where each of those goes. Until now the site named the issue tracker and nothing else, so an institution with a budget line and no GitHub account had no address at all.

One inbox rather than a form per page, because a second form is a second payload to keep in step and a second place a message can go unread. The message reaches the CRM that already fans a lead out to the sender's confirmation mail, the team notification and Slack, so nothing new has to be built or watched. Nothing is stored here on purpose: a dispute is published beside the number it argues with, which is why /api/challenge writes to disk, and an enquiry is private, which is why this one does not.

Two support pages rather than one translated page, for the reason D69 gives. Funding is the most local thing on the site. The ground page names research grants and multilateral budgets in general; the Brazilian page names the windows a Brazilian institution actually holds, and that list is not a translation of anything.

Cost. The route depends on a service outside this repository, so the form fails when Core does, and the failure is only visible in the logs. The pages name funding venues that change, which is prose that will go stale and that no test catches. Adding support to LayerSectionId means every future layer has to answer whether it holds its own support page, even where the ground-layer one would do. The contact form is English on a page a Portuguese reader can reach; the Brazilian page says so and says the reply comes in Portuguese, which is a statement about how the team behaves rather than something the code enforces.

Overturned by. Messages arriving that the CRM cannot route, which would argue for a repository-side record the way disputes have one. A second country layer whose institutions want a support page identical to the ground-layer one, which would make the per-layer section the wrong shape and argue for one page with a country-specific funding block. Evidence that readers use the form to file objections about specific scores, which would mean the split between /contact and /challenge is legible to us and not to them.

D72: Agenda items carry navigation to related institutions

Recorded 2026-08-30. Extends D35 and D54.

Choice. Each dimension-level item in a generated country agenda carries an institutionIds array. The ids point to nodes in that country's data/institutions/{ISO3}.json network. The array is computed from the node's existing dimensions field, which is already the curated navigation link from an institution to an NCB question. Countries without an institutional network carry an empty array. The links are explanatory navigation only: they do not add evidence, alter a score or change confidence.

Why. The agenda and institutional map previously met only through a reader who knew both files. The new field makes the path from a current agenda item to the institutions worth investigating explicit, while keeping one mapping to curate. Brazil is the first test: its agenda now carries the ids of the institutions tagged for each of the nine dimensions.

Cost. A broad dimension can point to many institutions, especially in Brazil's scaffolded state layer. The ids say where to look, not which actor is responsible for a specific gap or that it delivered well. A future item-level relationship may need a more specific, sourced link when the agenda stops being dimension-level.

Overturned by. Evidence that the node dimension tags are too broad to help readers navigate, which would justify a separately curated agenda relationship with its own source and scope. A country layer whose agenda items are more specific than dimensions would justify adding stable item ids before extending the link.

D73: Navigation is one tree, resolved against the path, drawn as a breadcrumb and tabs

Recorded 2026-08-30. Supersedes the nav rules in D69 and folds in the method subnav.

Choice. The viewer has one navigation tree, in apps/web/src/lib/nav.ts, and one rule for what is current. navRows walks it against the current path and returns the rows to draw; HeaderNav draws them and nothing else in the viewer draws navigation. Every row sits in the header.

The tree is four deep and the walk stops there:

1. the sections 2. the country you are in 3. which reading of it, where the project has written more than one 4. the pages of that reading

It reaches the reader as two bands rather than four rows. Everything above the deepest level is a breadcrumb in the header; the deepest level is a tab strip on its own rule under the header. Stacking four nav rows failed a look test: three levels each drew the same lime underline for current, so nothing said which one was the deepest, and two adjacent rows read as one group. The breadcrumb and the tab bar are borrowed shapes, and the borrowing is the point. A reader already knows what a slash-separated trail means and what a tab does, so neither has to be learned.

A trail of one crumb repeats the section already marked in the row above it, so it does not render. Method and Capabilities go straight from their section to their tabs, and the breadcrumb appears only where there is a trail to state.

The markup is copied in rather than installed, the way Icon.tsx copies Lucide paths. Both are plain elements with Tailwind classes.

A node declares how it claims a path and what opens under it. Countries has 52 children and no control can show them, so its child is resolved from the path: the crumb names the one country you are in, never the set. Method and Capabilities enumerate their children, because nine fits in a tab strip.

A country layer is a reading of that country and sits beside the English one, never under its pages. That is what keeps the tree four deep rather than five, and it is also what the layer is: another way to read the same country. Both readings render in the same crumb, separated by a middot rather than a slash, because they are alternatives at one level and not steps in a trail. A country with one reading skips that level, so its trail ends at the country.

A crumb marks its selection only when it holds more than one node. A lone crumb is a step, and a step is an ancestor of the current page rather than the page itself, so it carries no marker. A crumb offering both readings marks the selected one with the same lime underline the sections and the tabs use, which keeps one meaning for that mark across the whole nav and stops the selection resting on a colour difference alone.

Because the rows render in the header, above every layout that could open a file, which surfaces a country has must be knowable without touching the filesystem. hasLocalDestination already read the layer registry; INSTITUTION_MAPS in apps/web/src/lib/layers.ts now names the countries whose institution map is published.

This replaces three mechanisms that each had their own idea of what counted as current: PRIMARY_NAV with primaryNavOwns, METHOD_SUBNAV with methodSubnavOwns, and CountrySubnav with countrySubnavOwns. The first two rendered in the header and the third in the page body, so the site had two places a reader had to look to find out where they were.

Why. Three ownership rules is three chances to disagree, and they did: a country page lit Countries in the header while its own row of pages appeared somewhere else entirely, and the layer arrived with a fourth rule. One tree makes the depth a property of the structure rather than of whichever component happened to render. It also forces the honest question at every node, which is what a level means: section, then which reading of which country, then which page. A level that cannot answer that does not belong in the nav.

Costs. The header carries a breadcrumb under the sections, and a tab strip sits beneath it, where a country's pages used to sit at the top of the page body. The tab strip is a second band with its own rule, so the site has two horizontal rules above the content instead of one. INSTITUTION_MAPS is a hand-maintained list that can fall out of step with data/institutions/*.json; a stale entry costs a link to a page that says the country is not mapped yet, which is a state that page already renders. Pages outside the seven sections, /support, /contact and the comparison's own deep links, light nothing at level one except where a node claims them, so a new top-level page has to be placed in the tree or it will read as orphaned.

Overturned by. A fifth level that cannot be collapsed into a reading or a page, which would mean the four-deep cap is wrong rather than the content. Evidence that readers do not read the middot crumb as a choice between readings, which would justify a labelled switcher instead. A country whose pages do not fit in one tab strip, which would justify the docs-site shape of moving the lower levels into a left rail.

D74: Brazil's institution map uses an explicit inclusion rule and a federative pilot

Recorded 2026-08-30. Extends D56 and D58.

Choice. An institution belongs in a country map when it is a federative-level entity that allocates authoritative public data on at least one NCB dimension indicator, or a federal entity that exercises authority over a capability-relevant activity at national scope and has a public source citation a Brazilian reader can follow in under two clicks. The rule is descriptive: inclusion does not measure performance, enter a score or imply that the institution delivered well.

Brazil's first federative pilot names the state finance and planning functions alongside the São Paulo State Public Defender: Sefaz-SP, Sefaz-MA, Seplag-MG, Sefa-PA and the São Paulo Defensoria Pública. The rest of the state layer stays scaffolded until its sectoral institutions have the same source-backed treatment.

Why. The original map showed the federal lens well but made the state layer look more complete than its evidence warranted. Naming the rule makes expansion repeatable, while the five-entry pilot tests whether the same concepts survive when authority, data and delivery sit below the Union.

Costs. A public source can make an institution easy to find without proving that it has capability in practice. The pilot also leaves many state functions outside the map, and the two-click test depends on websites that can move.

Overturned by. Evidence that the inclusion rule systematically favors bodies with better websites over bodies with greater authority, which would require a separate source-access criterion. A state-level map whose institutions cannot be described without adding a new level or system, which would overturn the current federative pilot shape.

D75: The thesis argues, the about page describes, and the front page carries one module per section

Recorded 2026-08-30.

Choice. Three public pages are split by job and never restate each other.

1. /thesis owns the argument. Why capability is worth measuring, what the claim is, and how far the current data supports it. It draws the wealth correlation for all nine dimensions, so the reader sees the result rather than a promise to look it up. The two provisional layers are described as computed and unpublished, gated by D65, never in the future tense. 2. /about owns the object. What the dataset is made of, how big it is, when it last ran, what the benchmark refuses to do, what it gets wrong and how to argue with it. It states no second version of the claim, and links to the thesis for it. 3. The front page carries one module per section of the site. Each module is a sentence, a live number and one link out. No module reprints a list that has a page of its own.

Numbers inside these pages are computed, never typed into the copy. readWealthTracking in apps/web/src/lib/wealth.ts is the only place the GDP correlation column is summarised, and the front page, the thesis and the about page all read it, so the three cannot state different results for the same run.

Why. The two pages had grown into one argument told twice: both listed the nine capabilities, both named Brazil as the first case, both named Envisioning, and both stated the wealth claim. Two statements of one claim drift apart the first time only one of them is edited, and the about page had already drifted into promising a diagnostic result that the diagnostics had answered several releases earlier. The front page had the opposite failure. It printed the nine capabilities twice, once as the histogram's switch and once as a card grid, and compressed every other section of the site into a single paragraph of inline links, so a reader could not tell that agendas, limits or the decision log existed.

Costs. A reader who lands on /about looking for the argument has to follow one link. The front page is longer, and each module has to be maintained against the section it stands for: a new top-level section needs a module or it goes unmentioned. The widestSpread sentence names whichever country currently has the widest gap, so it can change between runs, which is the point of computing it rather than writing it.

Overturned by. Evidence that readers arriving on the about page expect the argument there and do not follow the link would move the claim back. Evidence that the front page's modules are read as a menu rather than as the benchmark would return it to the chart alone. A front page that grows past one module per section is the signal that this rule has stopped holding.

D76: AI scouts the evidence queue, but source checks and publication stay separate

Recorded 2026-08-30. Extends D20, D31, D33 and D52.

Choice. Evidence expansion becomes an AI-first, source-first workflow with three explicit objects:

1. A research slot is an uncovered country × declared-gap combination, selected by a deterministic inventory of the current evidence corpus. 2. A research lead is an AI-generated hypothesis containing a route, a possible mechanism, a scale test, disqualifiers and source-search targets. 3. An evidence record remains the source-checked, human-approved object in data/evidence/records.json.

pnpm bench research inventory writes the queue and balance report. pnpm bench research scout uses the AI Gateway to generate bounded leads, and pnpm bench research critique red-teams them against the five inclusion tests. Runs are immutable artifacts under data/research/runs. They are validated but never loaded by scoring, confidence, the agenda generator or the viewer's documented-delivery table. No AI stage can approve publication.

The scout must distinguish a possible institutional case, a source-backed dataset task and a gap that should not be forced into a case story. It must return search targets rather than invented URLs, numbers or coverage claims. The researcher still opens the source, verifies the metric and reference period, writes the limits and pattern fields, and runs the existing evidence validation gate before adding a record.

Why. The current evidence layer has 52 records but is concentrated in 20 countries, 15 gap indicators and one dimension, while its own reversal quota is below target. Hand-authoring one story at a time will keep selecting visible countries and familiar programmes. A deterministic slot inventory gives the AI a bounded search space that points at the actual missing cells rather than at whatever country has the best English-language documentation.

AI is useful for breadth: it can generate competing search hypotheses, identify adjacent constructs, propose source families and attack a lead before research time is spent. It is unsafe as an unreviewed fact source. Separating leads from records preserves D20's score boundary and D33's selection discipline while making the discovery process repeatable and auditable.

The first planning milestone is 100 qualified deliveries across at least 40 countries and 20 gap indicators, with at least 20 reversals and no country above one-third of the corpus. This is a corpus target, not a scoring threshold.

Costs. AI scout output can still be plausible and wrong, so a valid JSON run is not evidence and a critique is not a source check. The extra stages add research artifacts and review time. The deterministic queue can encode the existing corpus's blind spots, so its priority is a starting order rather than a claim about which countries matter most. AI model choice, prompts and source access also become part of the research record and need versioning.

The current evidence schema does not gain AI-authored factual fields. That keeps the published contract stable and means an AI run cannot accidentally be served as a delivery. If future work needs source packets, reviewer identities or automatic draft comparison, those fields require a new decision rather than being smuggled into EvidenceRecord.

Overturned by. Evidence that AI-generated leads systematically reduce source quality or increase conceptual mismatches compared with unaided research, which would narrow AI to clerical extraction. Evidence that the research queue is too broad to review, which would require a smaller slot contract or a separate country-set research frame. A source-backed comparable series that closes a gap supersedes the case lane for that indicator under D20.

D77: Correct the D76 corpus-balance description

Recorded 2026-08-30. Clarifies D76; it does not change the workflow.

Choice. The current evidence corpus is represented across all nine dimensions. It is uneven rather than one-dimensional: Building has the largest share at 19 of 52 records, followed by Adaptability at 8, while the remaining dimensions have between 1 and 5 records. The AI research queue must therefore expand countries, gap indicators and thin dimensions, while preserving the reversal and concentration guardrails.

Why. D76's phrase "one dimension" was inaccurate. Leaving it in the decision log would give the AI queue the wrong diagnosis and make the research plan less auditable.

Overturned by. A refreshed inventory showing a different distribution supersedes these counts; the rule to derive the diagnosis from the inventory remains.

D78: Taking part is a section of the site, and one registry declares the ways in

Recorded 2026-08-30. Extends D71 and D75, and renames the surface D50 reads.

Choice. Four things, together.

1. Participate is a section. /support, /gaps, /objections and /contact sit under one node in apps/web/src/lib/nav.ts, and the footer column derives from it. Before this, /support and /contact were in no navigation at all: a reader who had decided to help could only find them through a sentence at the foot of another page. 2. One contribution registry. packages/core/src/model/contribute.ts holds every way in. Each entry names the ask, what it has to carry, who usually brings it, how much work it is and what changes in the published benchmark when it lands. ContributionList renders it and contributionHref in links.ts addresses it. No page writes its own list of ways to help. 3. The gaps are published as open work. /gaps is a second reading of ingest: 'gap' and ingest: 'retired' from the indicator registry, one entry per gap with what it tries to observe, why it is open and a link that opens the contact form on that indicator. No new data. 4. /challenge becomes /objections. The old address redirects permanently. The score-side control keeps the verb, because challenging a score is what the reader is doing there.

The funded pieces in FUNDABLE_PIECES each state a scope and an effect, so a reader can size the work before writing.

Why. The project asked for help in four places, in four vocabularies, and scaled every ask to an institution. Nothing a single reader could do in five minutes was addressed anywhere, while the most specific request the project can make, the 26 declared gaps, rendered only as rows in a registry table. The word challenge read as a competition to somebody arriving from outside, which is the opposite of an invitation to argue. A registry rather than four page sections means a new way in is declared once and appears everywhere it belongs, which is the rule the indicator registry and the glossary already follow.

Costs. The contribution registry is prose held in code, so a copy change is a build. /gaps restates registry rows that /indicators already renders, which is a second surface over one source and has to stay a second reading rather than becoming a second list. The rename leaves a redirect to carry, and older links and citations that name /challenge now take a hop.

Overturned by. Evidence that readers use the participate section as a menu and never act would mean the ways are wrong rather than badly addressed. Evidence that /gaps produces unusable dataset suggestions in volume would mean the five-minute rung needs a higher bar, not a wider door. If a country layer needs its own contribution registry, because its institutions take part in ways this one cannot express, the registry becomes per layer and this decision is superseded.

D79: Headings are short and plain, and Octa is the wordmark plus one title

Decision. Three rules on the rendered surface, two for language and one for type.

1. A heading states what the section holds, in about six words. It stays a statement that could be true or false, per the existing rule, but it drops the aphorism. "How much of this is income?" replaces "The claim is that capability is not simply wealth". "What each country should work on" replaces "Each country gets an agenda, not a grade". "Nine scores, read together" replaces "A country is a shape, and no country is one number". The "X, not Y" inversion goes out of headings entirely rather than being capped at one per page, because a heading is where the reader notices it. 2. Octa is the wordmark and the front page hero title. PageTitle renders Inter. HeroTitle in apps/web/src/components/ui.tsx is the only Octa title on the site and is one step larger than a page title, at text-4xl sm:text-6xl against text-3xl sm:text-4xl.

3. The project describes itself by what it measures. The footer line, the metadata description and the social card carried "measures what a country can do, apart from how rich it is" over "52 countries, nine capabilities, public data, no ranking". All three now name capabilities: "measures whether a country can anticipate change, coordinate around it and build what it decides to build". The refusals keep their home on /about, whose job is to hold them.

The same pass removed the connective filler that named two pages and joined them with "and", of which "About says what it is made of and what it refuses to do, and contact reaches a person" was the clearest case.

Three further rules came out of the same reading. Reader copy carries no internal fact: "one inbox for the whole project" is a contract between agents and a reader never asked how many inboxes exist. Reader copy never reassures: "a person reads it and replies" invites the suspicion it denies and promises what the page cannot keep. And three short sentences in a row is the rhythm of generated text whatever the words are, so vary the length or cut to one.

Why. The maxim heading was the project's house style and it read as generated text. The self-description had the matching fault: a benchmark that introduces itself as not-wealth and not-a-ranking has told the reader two things it is not before naming one thing it does, and a count is not a claim.

On the headings, every section opened with a small thesis, so a reader scanning the page met nine arguments and no map, and the pattern told them nothing about where to click. The heading rule that produced it, "a statement that could be true or false", is still right; what it lacked was a length and a subject. The subject of a heading is the section, never the project's belief about the section. Octa had the same problem in type: a display face on every page title is a body face with extra weight, and it stopped marking anything. One Octa title per site restores the signal, and the front page is where a first-time reader meets the name.

Costs. Some headings lost information that now has to sit in the hint line below, which is one more line to keep true. Short headings are more likely to collide across pages, so "Known failures" and "Datasets we rejected" now appear in more than one place and a reader arriving from search sees less context. Dropping Octa from page titles makes the site quieter, and the brand mark now rests on the wordmark, the accent and the layout rather than on the typeface.

Overturned by. Evidence that readers cannot tell sections apart from headings alone would mean the six-word target is too tight for this material. A brand revision at envisioning.com that puts Octa back on section titles would supersede the type half of this decision, which follows the parent system rather than setting it.

D80: The site has five sections, and a page joins the one that answers its reader's question

Recorded 2026-08-30. Amends the tree D73 declared and the placement D75 and D78 assumed.

Choice. The header carries five sections and no more:

`` Countries · Capabilities · Method · Participate · About ``

Two sections fold into others, and nothing about the pages changes.

1. Agendas moves under Countries. The row under Countries answers "which country". Once the path names one, that country is the row, because 52 will not fit in a control. Before it does, the row offers the readings that take the whole set: all countries, compare, agendas. COUNTRY_INDEX_PAGES holds them. An agenda index is a way of reading the country set, not a separate thing the site is about. 2. Thesis moves under About. ABOUT_PAGES is overview, thesis, changelog. D75 gave those two pages different jobs and that stands: the thesis argues and the overview describes. Both answer the question a newcomer asks before either of them, which is what a section is.

Adding a sixth section is a decision, not a page. A new page joins the list whose reader's question it answers, all four of which live in apps/web/src/lib/nav.ts beside the tree.

READING_PAGES in the same file flattens the destinations for /llms.txt, because a machine has no rows to open. The footer renders four columns from the same lists.

Why. Seven sections made the header a menu of the project's own internal distinctions. Thesis and About were separate because they do different jobs, which is a reason for two pages and not for two headings: the site was saying "here is what we are" twice at the top level, and a reader arriving cold had to guess which one to open first. Agendas was a section because it has an index page, which is not the same as being one of the things the site is about. A section is a reader's question, and there are five of them: which country, which capability, how is it built, how do I take part, what is this.

Costs. The thesis is the project's strongest single page and it now sits one click below the header. The agenda index loses its permanent slot, and inside a country the header no longer offers a route back to compare or to the agenda index, only to the section itself. Countries' second row changes membership with the path, which is one more rule than a fixed list, and it is why that behaviour is written down here and commented at the node.

Overturned by. Evidence that readers cannot find the thesis, or that traffic to it collapses against its old placement, would restore it as a section rather than reopen the grouping. Evidence that the changing row under Countries reads as instability, rather than as one question answered two ways, would split it into a fixed row plus a country crumb, which costs a level of depth D73 caps at four.

D81: The viewer owns one button, one field and one control, and no shadows

Recorded 2026-08-30. Extends the brand contract D18 and D79 set.

Choice. Four primitives in apps/web/src/components/ui.tsx hold every interactive and panel geometry on the site.

1. Button, ButtonLink and buttonClass. 40px tall, 32px at size="sm", a 12px medium label at both, rounded-md, a 200ms colour transition. Four variants carry emphasis and nothing else: default is the filled action, accent is lime on navy ink and a page carries at most one, outline sits beside content, ghost goes inside a control that already has an edge. 2. fieldClass. 40px, rounded-md, body type, because a reader writes prose into a form field and 12px is not a size to write in. A field and the button under it share a height, so a form lands on one rhythm. 3. controlClass. 32px, rounded-md, label type. A filter names a choice rather than taking prose, and it matches the small button so a filter strip mixing a select and a button sits on one line. 4. Card. rounded-lg over a 1px rule, three fills and three paddings.

Two consequences follow. The viewer publishes no shadows. envisioning.com specifies a four-tier neutral shadow scale and puts shadow-md on cards; this site uses none. Navigation carries label type at every level, so there is no text-sm anywhere, which the type scale in AGENTS.md already stated and the header alone was breaking.

The front page opens on hero-band, a flat dark section on the footer's surface. The band spans the window through full-bleed and carries its own container: main is a centred column, and a dark surface that stops at the column edge reads as a card sitting on the page rather than as the page opening. That is how envisioning.com draws every section, and it is why html and body now clip their horizontal overflow, which absorbs the scrollbar width 100vw overshoots by.

The viewer draws no gradient anywhere. The band's one decoration is the dot motif, which envisioning.com keeps in lib/design/dot-motif.ts: a small opaque centre inside a larger translucent bubble, at an ambient alpha of 0.15. NCB draws one still frame of it as an SVG pattern instead of the parent's canvas animation, because a research surface does not need a background that moves and a still pattern costs no JavaScript. DotField is atmosphere and carries no meaning: the same motif carrying meaning is FlagBubble inside FlagField, where the centre is a country and the ring is how well evidenced its score is. Those two must never be confused on one page, which is why the field is aria-hidden and lives only in the band.

Navigation hangs from one edge for the same reason. The sections sit right from md up, so the crumb trail and the tab strip do too.

Why. Before this, eight files hand-rolled a button and reached five different geometries: px-4 py-2, px-3 py-2, px-3 py-1.5, px-2 py-1 and px-1.5 py-0.5. Two files declared a CONTROL constant with different padding under the same name, and two more inlined a third copy. Nine files repeated one card class string verbatim. Every one of them was individually reasonable and together they meant the site had no button, so a change to how a button looks was a change to eight files and a chance to miss one.

The shadow omission is a real divergence from the parent brand and is recorded rather than left to look like an oversight. An elevation under a table of scores reads as chrome and competes with the score bands, which are the one thing on the page that has to be read by tone. The 1px rule already separates a panel from the page on both themes.

The nav size is the other divergence. envisioning.com's nav is 14px; NCB's header stacks sections, a crumb trail and a tab strip, and the crumbs and tabs were already 12px. Making the sections 12px too means a reader locates themselves by the lime underline and the position, not by a font size that only distinguished the top row.

What it costs. The three heights are three numbers to keep apart, and the next agent will reach for the wrong one at least once. A 12px section link is a small target, so the desktop row carries py-2 and the mobile sheet keeps py-3. Two panels in EvidenceList and CheckList still hold their classes by hand, because the :has() rules in globals.css select on them, and a reader of those two files will not find the Card the rest of the site uses.

Overturned by. Evidence that a 12px section link is missed or mis-tapped would take navigation back to 14px and record it as the scale's one exception. Evidence that panels do not separate from the page on a real screen, rather than in a design review, would bring back the shadow scale. A second dark band proving useful on a page other than the front page would turn hero-band from a one-off into a section treatment.

D82: The drawn network is a separate deployment reading a published feed

Recorded 2026-08-30. Extends D54, D56 and D59. Does not supersede D56: the ledger stays the authoritative reading of a relation.

Choice. The project publishes an explorer feed for each mapped country at data/out/institutions/{ISO3}.{lang}.json, written by pnpm bench institutions and served at /api/institutions/{ISO3}?lang=. The feed is a projection of data/institutions/{ISO3}.json and never a second source of truth: every institution, relation and source comes from that file, and every label comes from a lexicon.

The network is drawn by @envisioning/app, which is not a dependency of this repository. It is closed source and this repository is public, so a viewer that imported it could not be installed or built by anyone outside Envisioning. The explorer is therefore a separate deployment that reads the feed over HTTP.

localizeInstitutionNetwork in packages/core/src/i18n/institutions-pt-br.ts is the one place a country file's own prose is rendered into a language. The viewer's ledger and the feed both call it, so the two surfaces cannot describe the same institution in different words.

Only a jurisdiction the map marks baseline or pilot is worth drawing. DRAWABLE_COVERAGE reads that status from the file rather than from a hand-written list.

Why. D56 refused a node-link diagram and its reasons were specific: 13 relation verbs rendered identically, a star topology that spent the canvas on spokes, and labels below the type scale. Two of those are now answerable. The force layout has a selected-item mode that draws concentric rings around one institution, which is the interface D56 described in words, and its labels are measured pills at mn and sm rather than 9px SVG text.

The third is not answerable and the data says so louder than before. The Brazilian map has grown to 359 institutions and 378 relations, and 261 of those institutions carry exactly one relation. The 26 scaffold states are each one hub with about eight spokes: 26 copies of the same star. Drawing all of it would show a fringe, not a structure. The union carries 85 internal relations and Sao Paulo 22, and those two do have a shape, which is why the drawable rule reads coverage status.

A picture also cannot carry a relation's direction. RELATION_FAMILY_STRENGTH maps the four families onto line width, and that is the whole of what the drawing distinguishes. Direction and verb stay in the ledger, which is why the ledger remains authoritative and the network is a second reading beside it.

Keeping the drawing outside this repository is what makes it affordable. An in-repo mount would add about 30 transitive dependencies to a repository that today depends on Next, React and its own core, pin React to an exact patch version, and require a private registry token to install. The cost of the separate deployment is one HTTP hop and one published shape.

Cost. The feed is a published artefact that can go stale against the map behind it. pnpm bench institutions is a separate command rather than part of pnpm bench all, following velocity, leverage and residual, so a change to data/institutions/*.json that does not rerun it leaves the explorer showing the previous map.

Language is baked in at write time, because the feed embeds rendered labels. A new lexicon means a new file rather than a parameter, which is the same trade the agenda documents make.

The explorer's chrome is not this project's. It brings its own header, menu and URL grammar, so it cannot sit inside the navigation tree D73 and D80 describe. It is reached as a leaf from the institution page and is not a nav tab.

level still carries Brazilian assumptions, as D56 already recorded, and LEVEL_COLOURS now spends the one accent on federal. A country whose capability backbone is not federal will need that mapping revisited rather than reused.

Overturned by. Evidence from readers that they arrive wanting the whole network shape, which would make the drawing the institution page's first surface rather than a leaf beside it. Or a relation schema that records direction and scope in a way a line can carry, which would remove the reason the ledger stays authoritative. Or @envisioning/app becoming installable from a public registry, which would make the in-repo mount cost only its dependencies.

D83: V-Dem civil-society strength is the first adapter-backed Coordination repair

Recorded 2026-08-31. Extends D23, D55 and the Coordination queue in the research roadmap. Promotes civil_society_strength from a gap to an adapter.

Choice. Add V-Dem's country-year civil society participation index (v2x_cspart) from the pinned Country-Year Core v15 release as a third Coordination indicator. The adapter fetches the public CC BY-SA 4.0 archive, extracts the 2024 country-year rows and emits the existing observation shape; the release, variable and year are fixed in source-catalog.ts. All 52 benchmark countries are present, so no country is silently carried forward or excluded. The source remains explicitly expert-coded (expert_panel), and the confidence penalty for that tier stays visible beside every score.

The first rescore gives Coordination a third observed row for 51 of 52 countries (the coverage floor therefore publishes 51 scores). Coordination's Pearson correlation with log GDP per capita is 0.561 (Spearman 0.611, n=50), down from the 0.627 Trust correlation and below the 0.70 wealth-proxy screen. The new row's own correlation is 0.395 and its wealth-attribution delta is 0.046; it forms no redundant pair with an existing Coordination row. The row does not erase the remaining gaps: university-industry collaboration and public-private collaboration stay unmeasured, and the single-country score below the floor stays visibly null.

Why. The index measures autonomy, organisation and participatory reach of civil society rather than government reputation, so it answers a declared Coordination construct without restoring the four mutually redundant WGI composites retired by D23. It is not administrative performance data and it is not a substitute for cross-agency delivery. Publishing the expert-coded source with its tier and keeping the wealth diagnostics live is the smallest repair that adds a full-frame signal while preserving those limits.

Cost. Coordination scores move materially because a new 0–1 frame is added; old values are not comparable, so this is a minor dataset release and every generated output is restated. The adapter depends on the publisher's archive layout and on unzip; a future release must update the pinned variable or fixture deliberately rather than changing the URL in place. V-Dem's expert coding can still reflect the evaluators' assumptions, so a lower GDP correlation is not proof that the measure is unbiased. Court performance, cross-agency delivery and the Trust institutional family remain open work.

Overturned by. A release revision that changes the construct or removes country coverage below the half-frame gate; a diagnostic release that pushes the row above the wealth-proxy threshold or into a redundant pair; or direct validation showing that the CSV values cannot be reproduced from the pinned archive. Any of those would return the row to gap or replace it with a more observable series.

D84: An indicator is checked for whether it still separates countries

Recorded 2026-08-31. Extends D22 and D42. Adds discriminationTrend to data/out/diagnostics.json and a section to the report. Changes no score.

Choice. Report, for every scored indicator, the interquartile spread of its normalised values in each year of a twenty-year window, and whether that spread is narrowing. Two flags: fading, where the rank correlation of spread against year is at or below -0.5 and the spread has narrowed by at least a quarter, and lowDiscrimination, where the latest spread is under ten points regardless of what it has done over time. The window's rules are the trend layer's rules rather than new ones. Historical values are scored against the frame built from every country's current values, so a change in spread is a change in the countries and never a change in the scale, and the spread is computed on the countries observed at both ends, so an indicator that gained coverage does not read as one that gained variance. The balanced panel size, the filled year count, the carried-forward cell count and the clamped cell count are published beside every row, because each of them can hold a spread still rather than measure it.

Why. Every other block in the diagnostics is a cross-section of the latest year: how a series correlates with income, whether two series carry the same information, how much of a dimension rests on judgement. None of them can see a series that has stopped varying. An indicator whose cross-country spread has collapsed still contributes a full share to its dimension's equal-weight mean, and what it contributes after that is closer to noise than to a measurement. That failure is silent, it arrives gradually, and nothing in the pipeline would have noticed it. The first run finds it in two places, both of them plausible as real convergence rather than as error: internet use, which the world has genuinely levelled on, and income inequality, on a series sparse enough that the carried-forward count has to be read beside the trend.

The test also answers a question the registry could not otherwise ask. Eighteen of the thirty-six scored indicators cannot be tested at all, because they start after the window opens or stop before it closes. That is a finding about what this benchmark can watch over time, and it now has a number.

Cost. A falling spread has more than one cause and this reports only the number. Real convergence, a publisher's methodology change, a survey series repeating its last round between waves and a frame that clamps early values all produce the same shape, and only the carried-forward and clamped counts separate them, and only partially. The twenty-year window is a choice: it is long enough to see a slow narrowing and short enough that half the registry still reaches it, but an indicator whose decay is slower than the window will pass. The flag thresholds are conventions with no theory behind them. Nothing reads the block, by design: it changes no score, no confidence and no weighting, so an indicator that fades goes on carrying a full share until a separate decision retires or reweights it.

Overturned by. A demonstration that spread on the fixed frame is the wrong statistic, most plausibly because clamping compresses early years enough to invert the sign for a class of indicator; or a fading flag that survives its carried-forward and clamped counts, which would make the block a trigger for reweighting rather than a report, and would need a decision saying what a faded indicator's share becomes.

D85: A section opens its pages from the header, and the header travels with the reader

Recorded 2026-08-31. Extends D73, which stands: one tree, one ownership rule.

Choice. Every section in the header opens its own pages. The section is one control and not two: it is a link, hovering it opens what it holds, and a click always means the same thing, which is go there. The panel shows what the tab strip would otherwise show only once the reader is already inside. Below md there is no hover to open anything, so the sheet keeps a separate chevron and the pages open in place under the section.

The panel reads the same level of the same tree the tabs read, through sectionMenuEntries in apps/web/src/lib/nav.ts. A menu and a tab strip cannot disagree about what a section holds, because neither knows anything the tree did not tell it.

One node needed a second answer. Countries resolves its row from the path, so on a country page the row is that country, which is what the crumb is for. A menu is opened from anywhere, and answering "which country am I in" to a reader who asked "what is in Countries" is answering a question nobody asked. A node may now declare menuChildren, the row it offers when opened rather than when walked, and Countries declares the three cross-country readings. It is the only node that needs it, and the field exists so the exception is stated in the tree instead of being guessed at by the renderer.

The panel hangs directly under the link, and the offset above it belongs to the panel rather than being a gap: empty space between a trigger and what it opens is somewhere for the pointer to fall through on the way down. A menu closes when the pointer leaves both, after a beat long enough to survive a cut corner, and on Escape with focus back on the link, and on arriving at a new page. Escape is read from the document rather than from the panel, because a hover opens a menu without moving focus into it.

A keyboard has no hover. ArrowDown from a section link opens its panel and lands on the first item, the arrow keys walk it and wrap at both ends, and Home and End reach the ends directly, because role="menu" promises a reader that they will. The item focused on a keyboard open is chosen in a render and not in a frame callback: a tab that is not painting never runs a frame, and the focus was silently lost there before this was written down.

The lime mark moved from the control onto the label inside it. The control now holds a chevron as well, and an underline drawn under a chevron is a smudge.

The header is sticky, and the tab strip rides inside it so the two bands stick as one. A menu the page can scroll out from under is a menu that has to be scrolled back to. html, body { overflow-x: clip } in globals.css is what allows this, and D81 chose clip over hidden for exactly this reason. The tab strip moved inside the header element rather than beside it, which also keeps the embed rule in globals.css pointing at one element to hide.

The sheet carries its own scroll at 70dvh. A sticky box taller than the window has a bottom that no amount of scrolling reaches, and the sheet with a section expanded is taller than a phone.

Why. The tab strip only exists once the reader is standing in the section that owns it, so the site's shape was legible only from inside. Nine method pages, three cross-country readings and four ways to take part were one click away from a reader who already knew they existed and invisible to one who did not. The footer listed them, which is a fine second answer and a poor first one. This costs the header a duplicate of the row already under it while the reader is inside a section, and that duplicate is the price of the same set being reachable from outside it.

Cost. Hover is not an input every reader has. A touch reader taps a section and goes there, which is what the tap did before this, so the panels are a desktop affordance and the sheet's chevron is what stands in for them. The sticky band is 189px on a country page, where the crumb takes a line of its own, against 120px on a method page and 72px where a section has no pages under it; on a short window that is a real share of the fold. A panel that opens on hover also opens on the way past, which is the price of not making the reader click for it.

Overturned by. Evidence that readers open sections they are already in more than sections they are not, which would make the menus a duplicate of the tabs rather than a way past them; or a header that has to grow a fourth row, which would make the sticky band too tall to keep.

D86: The front page states the claim under test, not the shape of the data

Recorded 2026-08-31. Divides with D75 and D80.

Choice. The dark band that opens the front page says what the benchmark is testing: that a country's capability is separate from its wealth, and that public data can see the difference. It no longer says how the data is shaped. The count of dimensions, the common scale and the confidence beside each score left the band, and the sections under it keep all three.

Three pages now hold three different things about one project, and none of them repeats another. /thesis argues why capability is the bottleneck now, which is a claim about the world. The front page states the claim the benchmark can fail, which is a claim about the instrument. /about says what the instrument is made of. D75 split the second and third; this splits the first from both.

Why. The band read as a specification. "Nine capabilities, scored from public data" answers a question nobody has asked yet: a reader who has not been told why capability is worth separating from income has no use for how many parts it has, and the two sections directly under the band already say it better, with the data drawn beside them. An opening that describes its own format spends the one place on the site where a stranger is still deciding whether the project is about anything.

Stating the claim also states the failure. "Tests whether" is a promise that the answer could be no, and the diagnostics section further down the same page reports how far the separation actually holds. An opening that can be wrong is worth more than one that cannot.

Cost. The band no longer says what the reader is looking at, so the first concrete fact about the data now arrives one section down. A reader who came for the numbers reads a sentence of argument first. The claim is also stated in docs/WHY.md and argued on /thesis, so three surfaces now carry a version of it and can drift apart; the front page holds the shortest one on purpose, and it is the one to change last.

Overturned by. Evidence that readers arrive already knowing why capability is worth measuring and leave because the front page will not tell them what it holds, which would make the band the wrong place for an argument.

D87: The Countries menu draws the country you are standing in

Recorded 2026-08-31. Extends D85.

Choice. The Countries menu opens with the shape of the country the path names, drawn as the same nine-axis radar the countries grid gives every card, above the three cross-country rows. Away from a country there is no shape and the menu is what it was.

This is the answer to a problem D85 created. That decision made the menu's rows path-independent on purpose: a menu is opened from anywhere, so it offers the readings that hold from anywhere, and a reader standing in Chile who opens Countries sees the same three rows as a reader standing in Japan. Correct, and it leaves the menu with nothing that says where the reader is. The shape says it, in the one language this site is built to be read in. It is also the argument for publishing no single number, made in the smallest space the site has: nine axes at a glance, and no total anywhere on them.

A node declares menuPreview and the renderer decides how to draw it, the same division menuChildren uses. The tree says there is a picture and what it is of; it does not know what a radar is.

The shape is fetched when the menu opens, from /api/shape/[iso3], which returns nine scores and nine confidences and nothing else. The header renders above every page, and the slim index is 597KB with no cache in loadIndex, so reading it in the root layout would make a reader on the glossary page parse every country in the benchmark to render a chart they will never open. The cache holds the request rather than the answer, so the second mount joins the first instead of asking again. The radar is loaded with the menu through next/dynamic for the same reason: it is the largest component in the viewer and it should not sit in the first bundle of a reader who never opens a menu.

The picture is not a link. The rows below it are where the menu goes, and the country's own profile is one click away in the crumb beside it, so a third target here would compete with both.

Why. A shape is what this project has instead of a rank, and the menu was the one place a reader could open, on a country page, and be told nothing about that country. It also costs nothing to a reader who does not open it, which is what made it worth adding to a header whose restraint is deliberate.

Cost. A chart that arrives over the network arrives late: the panel opens at its full height with an empty box where the shape will be, and on a slow connection the reader sees the rows before the picture. Holding the height is the deliberate half of that, because a panel that grows under a moving pointer is worse than one that waits. It is also the first thing in the viewer that draws data outside a page, which means the header can now be wrong about the data in a way it never could before: a shape cached at the top of a session survives a re-ingest that changes it.

Overturned by. A measurement that readers open the Countries menu to navigate and are slowed by a picture in the way; or a second menu wanting a preview of its own, which would make menuPreview a general mechanism rather than the one exception it is written as.

D88: The footer is the tree, not a copy of it

Recorded 2026-08-31. Completes D73.

Choice. FOOTER_NAV_GROUPS is derived from NAV_TREE: one column per section, labelled by the section, holding what that section opens. It was a hand-written list of four groups that spread the same page arrays the tree walks, close enough to look derived and not derived at all.

It had already drifted. Capabilities opened nine dimensions in the header and was a single link in the footer, and the footer's first column was called Explore, which is not a section and answers to nothing in the tree. A sixth section would have reached the header, the crumb, the tab strip and the sheet, and never the footer, and nothing would have failed.

The column count is derived as well. It was lg:grid-cols-5 in the markup beside a comment naming the one group that took two of them, which is the same kind of fact written down in the same wrong place: the moment the groups came from the tree, that number put a section on a row of its own. A group longer than six entries takes two columns and splits inside them, and the row is as wide as the groups need. The rule is the length and never the name.

D73 said one tree, one ownership rule, one renderer, and it meant the header. The footer was outside that sentence and stayed a copy for as long as the sentence stopped where it did.

Why. Two lists of one thing agree only while somebody edits both, and the proof that nobody had was sitting in the footer: a section that opened nine pages upstairs offered one downstairs. A missing footer link is the quietest possible failure, because the page it points at still exists and still works, and nothing anywhere reports that the way in went away.

Cost. The footer now shows every capability, which is nine more links in a place that carried 19 and now carries 28, and the sections no longer get to disagree with it on purpose. A footer that wants to say less than the header says, which is a reasonable thing for a footer to want, now has to say it in the tree or not at all. The width map spells out the classes Tailwind can see, so a tree wider than the map wraps to a second row rather than overflowing: that is the right failure, and it is still a failure that only shows up on screen.

Overturned by. A section whose footer column is the wrong shape for what it holds, most plausibly Countries once there is a way to offer more than the three cross-country readings; or a decision that the footer's job is a chosen short list rather than the whole tree, which would want the choice recorded in the tree as a field, never as a second array.

D89: The subnational layer is a computable diagnostic registry, not a second score

Recorded 2026-08-31. Supersedes the fixture language in D66 while preserving its national-score boundary.

Choice. Subnational output is published through a registry of country, geometry, source, unit, reconciliation rule and denominator method. Each published file carries the national value beside its constituent values, the retrieval metadata and a check block containing the recomposed value, the national value, the signed residual and a tolerance. aggregate is a claim that the build tests: validation fails when the recomposed value differs from the published national value by more than that series' tolerance. equal and population describe how the diagnostic recomposition is calculated; none means no recomposition is claimed. The denominator is therefore a calculation method, not evidence that the statistic is additively decomposable.

The 2024 IBGE state Gini fixture is restated as independent with an equal-unit diagnostic. Its published state mean is 0.486 against a national coefficient of 0.503, a signed residual of about -0.017. That residual is shown because it reveals the between-state component that a state range hides; it is not a failed build and it is not a national score. Population denominators are recorded from IBGE's 1 July 2024 federative-unit estimates so a future aggregate series can be recomposed without an unsourced weighting.

The registry and its files remain outside buildFrame, score, confidence, momentum, agendas and rankings. A subnational unit receives no capability score and no benchmark confidence. The first maintained geometry is state; publishing municipal output requires a separate decision about the unit of analysis and the resulting reader contract.

Why. D66 correctly kept subnational observations beside the comparison layer, but a hand-written reconciliation label let the first fixture assert a false aggregate relationship. Making the calculation and its inputs explicit turns the label into an auditable claim. A registry prevents every new series from growing a bespoke adapter, while a small index makes the research layer discoverable without filesystem probing. Keeping the Gini independent preserves the useful local variation instead of forcing a mathematically invalid national recomposition. The first candidate for expansion is SICONFI budget execution, but it must pass the source-memo and exact-variable checks before publication.

Costs. The output contract and path are breaking for consumers of the pilot, so the dataset version moves major even though national scores do not move. A per-series tolerance needs editorial review, and population estimates may not be the right denominator for every administrative statistic. The registry still needs one adapter per publisher family, and a state census can show variation without proving delivery quality. Municipal data is deliberately left unresolved rather than smuggled in through a state-shaped interface.

Overturned by. A reproducible decomposition method showing that the Gini fixture's state values can validly compose the national coefficient; evidence that readers mistake the diagnostic residual for a score; or a source release whose unit definitions cannot be mapped to the registry without silent imputation would require revising the contract. A successful municipal pilot with a clear audience and stable unit identifiers would justify a separate geometry decision, not a relaxation of the national boundary.

D90: Evidence coverage is a complete country-by-capability matrix

Recorded 2026-08-31. Extends D20, D35, D46 and D76.

Choice. The agenda index derives one complete matrix from the published evidence register: every country in COUNTRY_ISO3 is a row and every dimension in DIMENSIONS is a column, including the empty cells. A filled cell prints the number of documented deliveries filed against that country's gap indicators in that dimension and links to the same filtered delivery register D46 defines. Row totals state both capabilities covered and records filed; column totals state countries represented and records filed.

buildAgendaEvidenceCoverage in packages/core/src/pipeline/agenda.ts is the only aggregation. It reads each record's dimension through the indicator registry, so a record cannot be filed into a second hand-written capability map. The matrix remains a reading of data/evidence/records.json: it enters no score, confidence, coverage floor or research-run approval path.

An empty cell is labelled as an uncovered research slot, not as evidence that the country lacks the capability. A filled cell is evidence that at least one qualifying delivery was documented, not evidence that the whole capability is present. Record counts are therefore inventory counts and never capability weights.

Why. The delivery register could answer which cases exist after filtering, and the research inventory counted countries and dimensions separately, but neither showed their intersection. A reader could see 79 records and 22 represented countries without seeing that the current corpus fills only 44 of 468 country-dimension cells. Keeping every zero visible makes language bias, country concentration and thin dimensions concrete before the next source search begins.

The matrix also replaces the agenda index's separate country list. Every row still reaches the country's agenda, while the page no longer prints the same country set twice.

Costs. Fifty-two rows and nine capability columns require horizontal scrolling on a narrow screen. A count can conceal that several deliveries use the same publisher or bear on the same gap indicator, so the cell is a doorway to the records rather than a quality grade. As the corpus grows, large counts may become less useful than a status or source-family split; adding either must still preserve the one-record-one-delivery rule and the empty cells.

Overturned by. Evidence that researchers choose better-balanced source checks from a smaller queue view, or that readers consistently mistake filled cells for scored capability, would move the matrix off the public agenda index and leave it as a research artifact. A comparable series that closes a gap removes that indicator from the case lane under D20; the matrix must then be rebuilt from the remaining admissible records rather than preserving a stale cell for visual completeness.

D91: Country context and page links share one subnavigation band

Recorded 2026-08-31. Supersedes the visible rendering half of D73; the single navigation tree and path-based ownership remain.

Choice. The five global sections remain the only top-level row. Every route with child pages gets one subnavigation band beneath it. On a country route, the country context and the deepest page set share that band; on a layered country route, the country context, the English/local reading choices and the current reading's page set share it as well. The global section is not repeated inside the breadcrumb because it is already active in the row above. The tree in nav.ts still carries the levels needed to resolve ownership and links, but the renderer never gives each level its own horizontal row.

Why. The country screenshot made the navigation read as three steps: section, country and page. The repeated Countries label added no wayfinding, and a separate country row followed by a page row made the header taller than the page hierarchy required. Keeping the country context beside its page links preserves the reader's location and every destination while making the one subnav rule consistent with Method, Capabilities and the country index.

Costs. A narrow viewport can wrap the context and page links within the same band, so it is a flexible row rather than a guaranteed single line. Layered countries still have more choices in that row because their reading is a real alternative, and the breadcrumb remains a compact context selector rather than a full linear history. The navigation tree is deeper than its visible bands, which must stay clear in comments and accessibility labels.

Overturned by. Evidence that readers cannot tell country context from page links in the shared band, or that the wrapped row becomes harder to operate than the previous stacked treatment, would justify restoring separate context and page bands. A new kind of country layer with more than one reading or a page set that cannot fit the shared row would require a new layout decision rather than an exception in a component.

D92: The agenda history reaches for the oldest fetched year, but never invents a trend

Recorded 2026-08-31. Extends D22 and D24.

Choice. World Bank ingestion now requests history from 1976, which is the current 50-year look-back. The default momentum spans are ten years, twenty years, thirty years and the oldest configured ingest year. The last span is derived from INGEST_FROM_YEAR and the scoring year rather than hard-coded, so the horizon follows the source window. A span is published only when its matched basket still has at least two indicators and covers half of the country's current observed indicators in that dimension. The agenda chart offers the spans present in that country's published output, one dimension at a time, on today's 0–100 frame.

The older fetch added 3,611 observations without restating or dropping an existing value. In the current 52-country run, a 50-year matched basket exists for eight country-dimension cells; Brazil's dimension-level history reaches thirty years in Agency, Adaptability and Building, while individual indicators can reach further. A missing 50-year dimension line is therefore an evidence boundary, not a claim that the country did not change.

Why. The request for a longer country timeline is useful, and the raw World Bank series contain some pre-1990 history. Extending the fetch gives the chart a chance to show a genuine long movement while preserving D22's two guards against dataset churn: one current ruler and one basket held constant between the endpoints. Keeping ten and twenty years beside the longer views also shows when a long line depends on a narrower set of measures.

Costs. The observation and country output files grow, and a 50-year span is available for only a small minority of country-dimension cells because many indicators began later or are periodic. The current frame makes historical values comparable but not absolute development levels; values outside it still clamp and are counted. A 1976 start is a retrieval boundary, not evidence that every country is observed continuously from that year. The chart must continue to show gaps rather than fill them.

Overturned by. A source window with stable, comparable pre-1976 coverage would justify moving the ingest start earlier. Evidence that a longer matched trend systematically misleads readers about a discontinuity, or that its minimum basket hides more than it reveals, would justify removing the oldest default span while retaining the indicator-level histories.

D93: The history window reaches 1960, and the fetch must not truncate it

Recorded 2026-08-31. Supersedes the 1976 retrieval boundary in D92 while preserving its matched-basket and no-imputation rules.

Choice. INGEST_FROM_YEAR moves to 1960, the earliest year the registered World Bank request can usefully ask for across the benchmark's current source families. The default momentum spans keep ten, twenty, thirty and fifty years, then add the oldest configured window, which is derived from the scoring year and the ingest start. The World Bank request page size rises from 3,000 to 10,000 rows so 52 countries over the longer window arrive in one complete response. The ingest report remains authoritative: a source failure is carried forward, while a successful response is never silently truncated.

The first corrected 1960 fetch added 4,570 observations and dropped none. The previous incomplete attempt had recorded 4,251 apparent drops before the page size was corrected; the following run restored them as additions, leaving the observation file complete for the successful response. No value was restated.

Why. More history gives the agenda a wider context and lets the benchmark test whether a dimension's movement survives a longer comparison. The window is intentionally wider than the requested fifty years so a source series that starts in 1960 can be used at its earliest point. The published dimension trend still requires two indicators at both ends and half of the current observed basket. A missing long span is evidence coverage, not a zero and not an extrapolation.

Costs. Many World Bank indicators begin after 1960, so dimension-level trends will remain much shorter than individual indicator lines. The files and revision log grow, and the ingest boundary may span political or statistical breaks that the model cannot identify. A large page size is safe only while the country set and requested years fit below it; if they grow beyond that, the fetcher needs explicit pagination rather than another silent increase.

Overturned by. A registered source release with earlier, comparable data would justify moving the window again. Evidence of incomplete responses under the current page size, or a reader test showing that the extra context is mistaken for an absolute development score, would require pagination or a different presentation boundary before extending the range further.

D94: Dated agenda evidence anchors a separate history rail

Recorded 2026-08-31. Extends D20, D22, D35 and D92.

Choice. Every country's agenda evidence record carries its sourced started year into ownEvidence. The agenda history chart filters those records to the selected dimension and plots them by year on a separate rail aligned to the full configured source window. The capability line remains the matched indicator basket and retains its own span, base score, current score, confidence and clamp caveat. The event rail is context only: it never enters a score, confidence, momentum basket, agenda kind or ranking. Each marker has the record title and year in its accessible label, and the visible list below the rail keeps the items discoverable without requiring hover.

Why. The benchmark already records institutional changes beside missing indicators, but the prose list made it hard to see sequence and overlap. A separate rail puts a documented delivery next to the period in which the quantitative evidence moved without pretending that the delivery caused that movement. Keeping it separate also allows an older agenda item to remain useful when no dimension-level basket reaches back to its year.

Costs. A dated delivery is not a complete history and its start year is not the year its effects appeared. Several items can share a year, so markers may stack into lanes. The rail can show only evidence records with a sourced start year inside the configured window; earlier records stay in the agenda list.

Overturned by. Evidence that readers infer causation from proximity alone, or that the rail becomes too dense to identify individual records, would justify removing it or limiting it to an explicit event-selection view. A comparable series that directly measures the documented delivery could move it into the capability line under D20, but it would still not make the event marker a score.

D95: The country shape belongs to the field hover, not the Countries menu

Recorded 2026-08-31. Supersedes D87.

Choice. The homepage's FlagHistogram passes each country's nine scores and confidences into the shared FlagFieldPoint, and the field's existing hover card draws the country radar beside its one-capability position. The Countries menu keeps only the navigation rows it opens; it carries no country preview.

The radar is passive and not a second target: the flag remains the link to the country profile. It is loaded dynamically when a hover card needs it, so the field's initial bundle and the other FlagField surfaces do not pay for a chart that has not been requested. Other surfaces can opt in by supplying the same optional shape field, while the shared field geometry and hover behavior remain the only distribution implementation.

D87 put the right picture in the wrong place. A reader looking at the field is already asking what the country at that position looks like across the other capabilities; a navigation menu is answering where its links go. Keeping the shape with the flag makes the explanation arrive at the point of attention and leaves the menu legible as a menu.

Why. The homepage is the comparative entry point and already has the full slim profile for every country. Re-fetching one profile from a header menu added latency, a second data path and a large visual interruption to a control whose job is navigation. The hover card has the score, confidence, trend and within-country spread already, so the radar completes that reading without introducing a new interaction target.

Cost. The shape is available only while a flag's hover card is open and is not present in the Countries menu. A compact radar adds height to that card and can extend above or below the field near an edge, so the card remains pointer-transparent and the flag remains the only link. The homepage already ships the nine-value profile needed to draw it; dynamic loading limits the additional JavaScript cost until the reader asks for the shape.

Overturned by. Evidence that readers need country context while navigating from a country page, or that the homepage hover card becomes too tall to read, would justify revisiting the placement. A second surface needing the same shape should extend FlagFieldPoint or a shared profile component, not recreate the radar or add another menu-specific data path.

D96: Native tooltips are for exceptions, not metrics

Recorded 2026-08-31. Extends D18, D53 and D67.

Choice. The viewer does not attach native browser tooltips to every score, confidence meter, trend mark, table cell or chart point. Score bands are explained in the visible score legend; confidence bands are explained in the visible confidence legend; trend context is printed beside the change; and compact states such as clamping, missing coverage and rejected inputs are named in the row or a screen-reader-only description. SVG charts use their chart-level accessible names and visible event lists rather than <title> elements that appear on pointer hover.

The deliberate FlagField country card remains the one richer hover reading. It is part of the field's interaction contract, is pointer-transparent, and does not add a second link or a tooltip to each metric inside the card. See D67 and D95.

Why. Native title hovers are browser-controlled, inconsistent across devices, unavailable on touch, and easy to trigger accidentally when a table is dense. The same explanations were already present in page copy, legends, dialog content or linked detail pages, so repeating them on every number made the interface feel noisy without adding a reliable reading path. Moving the important meanings into the page also preserves them for keyboard and touch readers.

Cost. A reader no longer gets an extra sentence by pausing over an individual score or chart point. The replacement is intentionally less per-cell: legends and row labels explain the shared rule, while the field card is reserved for the chart that needs point-of-attention context. A future detail interaction should be an explicit, keyboard-reachable disclosure rather than another native title attribute.

Overturned by. Evidence that a specific dense view loses an important meaning after the change would justify adding a visible or keyboard-reachable detail control for that view. It would not justify restoring a repeated native tooltip to a shared metric component.

D97: Challenge entry is centralized in the header

Recorded 2026-08-31. Supersedes the score-side trigger in D62 and the score-side wording in D78.

Choice. The score-side Challenge control is removed from radar readouts and score tables. One Challenge action lives in the global header on desktop and mobile. It opens the public dispute form with country and capability selectors; the submission endpoint still validates the target and captures the current score and confidence from the country file. Contested badges remain beside affected scores because they report state rather than offer an action.

/objections remains the public ledger and explanation of how to object. The header action is the single entry point for a new score dispute, so a reader does not have to find a particular table cell before deciding to argue with the benchmark.

Why. A challenge button in every score cell repeated the same action across the radar, country profile, capability table and all-country table. That made the benchmark's invitation to argue compete with the scores themselves and made the most useful action depend on which surface happened to be open. A single header action is discoverable everywhere while the selectors preserve the target snapshot that D62 requires.

Cost. The form no longer inherits a target automatically from the number beside it. The reader must choose the country and capability, and the form does not show the selected score until the server records the submission. The public ledger and the contested state remain separate from the entry action so that removing a button does not remove the review trail.

Overturned by. Evidence that readers cannot identify the target from the selectors, or that objections fall because the score is no longer visible at the moment the form opens, would justify adding a target-aware preview to the central form. It would not justify restoring a button to every score cell.

D98: Confidence values use one compact display

Recorded 2026-08-31. Supersedes the confidence-meter presentation in D17 and the ConfidenceBar presentation described in D18.

Choice. Every viewer surface renders a confidence value as one band-colored chip containing its exact two-decimal number. The separate track and number are removed. Confidence remains a separate claim from the score; the visible confidence legend names the bands, and radar edges and flag rings continue to provide the chart-specific evidence encoding.

Why. A track followed by a number made one confidence value read as two controls and consumed more space than the score it qualified. One compact display keeps the exact value legible, preserves the band distinction, and lets tables, cards and prose use the same treatment.

Cost. The chip no longer uses bar length to show the magnitude within a band. The number is the precise magnitude and the legend explains the shared band colors; charts retain their own dashed-edge and ring treatments.

Overturned by. Evidence that readers mistake the chip's band color or lose the exact value would justify a more explicit single-element treatment, such as a chip that includes the band label, but not a return to a repeated track and number pair.

D99: Portugal joins the benchmark frame as a source-backed country

Recorded 2026-08-31. Extends D47, D52, D64 and D83.

Choice. Portugal (PRT, flag code PT) joins the single country registry. No indicator, weight or transform changes. The 53-country set therefore builds the normalization frame, and the published dataset moves from 5.1.0 to 6.0.0. The existing World Bank ingest, Joint EVS/WVS trust adapter and V-Dem adapter are rerun against the expanded set; no manual or synthetic Portugal values are added.

The source-backed release emits 841 World Bank observations for Portugal, one 2022 Joint EVS/WVS A165 trust observation and one 2024 V-Dem civil-society observation. Portugal publishes all nine dimension scores in the rebased output, with the lowest current scores in Experimentation (30.3) and Building (33.9). Trust is readable but remains very thin at 36.8 with confidence 0.159, so the profile is a measured research case rather than a claim of a complete national capability picture.

Portugal has no documented delivery records yet. The refreshed research inventory keeps it in the queue with 25 uncovered country-gap slots; those slots are leads for source and case research, not evidence and not scores.

Why. Portugal adds a useful Southern European comparison to Spain and the other European cases while preserving the benchmark's one shared frame. The country is added as a data-set expansion because it is an explicit research case, not as a side effect of closing an indicator gap.

Cost. Every score and historical normalized series is restated under 6.0.0; 5.1.0 values are not comparable to the new frame. The country addition also increases the current evidence matrix from 468 to 477 country-dimension cells. Portugal remains absent from the Brazil-specific country layer under D69, and its research queue remains empty of documented deliveries until a source-backed or evidence-track contribution passes the existing gates.

Overturned by. Evidence that Portugal is outside the benchmark's declared comparison purpose, that its source mappings are not reproducible, or that the expanded frame distorts well-evidenced countries enough to invalidate the relative scale would justify removing or revising this case in a subsequent versioned rebase.

D100: A retired row no longer lowers coverage

Recorded 2026-08-31. Amends D23 and D45; the confidence definition comes from D37's contract.

Choice. confidence = coverage x recency x source_quality keeps its shape, and coverage changes its denominator. It was observed indicators over every row the dimension declares. It is now observed indicators over the rows that could carry a number: the observed rows plus the declared gaps. Rows marked ingest: 'retired' leave the denominator.

Retired rows do not leave anything else. They stay in the registry, they stay published on each country's indicator list with status: 'retired', they stay counted in measurability in the diagnostics, and the evidence for rejecting each one stays in its registry note and its decision. A reader can still see that Trust declares eight constructs and measures two.

Why. A gap and a retirement are different facts about the world and the denominator was reading them as one. A gap says nobody publishes a comparable series, so the dimension cannot be measured here yet. A retirement says a series exists, this project inspected it, and it measures the wrong thing: the four WGI series correlating with each other between 0.93 and 0.98 and with log GDP per capita at 0.91 (D23, A4), the LPI recording how a country looks to freight forwarders (A9). Charging a dimension for refusing those was charging it for the discipline that makes the rest of the number worth reading.

It also put a ceiling nothing could lift. Coordination and Trust each declare eight rows and retire three. Under the old denominator their coverage could not exceed 0.625 however much data landed, which capped their confidence near 0.53 and made the good band unreachable by construction rather than by evidence. A confidence scale that a dimension cannot climb is not measuring how much is known about it.

Cost. Every published confidence value for Coordination, Trust, Building and Shared Purpose restates upward, and two dimensions cross a band boundary: Coordination from very thin to thin, Shared Purpose from very thin to thin. No score moves, the normalization frame is untouched, and no country is added or removed. The change makes four dimensions look better evidenced without a single new observation, which is real: the reader must still read observedIndicators beside the confidence, and A12 remains the entry that says what Trust and Coordination actually rest on.

The bump is minor. The frame holds and no published field is added or removed, so D37's major test is not met, but a consumer pinned to a confidence band will see cells move. D37's three cases do not describe a change to how a published number is computed, and this entry records that the bump was chosen rather than derived.

The change is not sufficient for its own motivation. Measured against the current output it moves the site mean confidence from 0.386 to 0.422, and Trust stays very thin at 0.209 because Trust's problem is two observed rows, not three retired ones. Coverage still has to be earned.

Overturned by. Evidence that readers take the higher confidence as a claim about evidence rather than about declared measurability, or a decision to retire rows for reasons other than rejecting an inspected dataset, which would make the denominator selective in a way this entry does not defend.

D101: Portugal's first evidence batch stays outside the score

Recorded 2026-08-31. Amends the research-state statements in D99 and extends D20 and D33.

Choice. Five source-backed records for Portugal enter data/evidence/records.json: national digital identity, the CoLAB applied-research network, the Qualifica adult-learning programme, Portugal Ventures, and the dismantled Novas Oportunidades programme. They document delivered institutional cases against four gap indicators, include one reversal, and remain evidence rather than observations: they do not enter a score, confidence value, normalization frame or dimension coverage count.

The refreshed research inventory therefore reports 204 deliveries across all 53 countries and 21 gap indicators, including five Portugal records. Portugal still has 21 uncovered country-gap slots, which remain the queue for comparable source work. The records' claims, dates, publishers and limits stay in the evidence ledger; no research lead is treated as a scored series.

Why. Portugal's source-backed country profile benefits from documented cases that show how institutions deliver, rebuild or scale capability, especially where the benchmark's comparable indicators remain gaps. Keeping the cases in the evidence layer preserves the distinction between a sourced national example and a country-comparable measurement.

Cost. D99's statements that Portugal had no documented delivery records and 25 uncovered country-gap slots are superseded by this batch. The dataset remains 6.0.0, all nine Portugal scores and the 53-country frame are unchanged, and the four gap indicators remain gaps until a comparable series covers at least two countries. If a gap is later promoted, these records may be stranded under D20; the decision and git history keep that research trail auditable.

Overturned by. A source correction, failure of the evidence inclusion tests, or proof that a record is an announcement rather than a delivered institutional case would justify removing or rewriting that record in a later evidence change.

D102: The confidence-method update keeps Portugal on the same frame

Recorded 2026-08-31. Extends D37, D99, D100 and D101.

Choice. The Portugal country addition remains the Dataset 6.0.0 frame rebase. The later retired-row denominator change in D100 is published as Dataset 6.1.0: it restates confidence values, but does not add or remove a country, alter the normalization frame or change any dimension score. Portugal remains one of 53 countries with nine published scores, five non-scored evidence records and 21 uncovered research slots.

Why. The semantic version separates the material country-set rebase from a method change that affects confidence but leaves scores and the frame intact. Naming both releases lets a reader distinguish Portugal's admission from the later confidence restatement while keeping the evidence and research trail continuous.

Cost. A consumer pinned to Dataset 6.0.0 will see different confidence values than one reading 6.1.0, even though the Portugal scores and frame are identical. The current Portugal Trust confidence is 0.255 under the 6.1.0 denominator, and its low confidence still limits how strongly that score should be read.

Overturned by. Evidence that the denominator change altered a score or the normalization frame, or that the version boundary hides a published-shape change, would justify revising the release classification in a later decision.

D103: Every chart draws from one set of weights and shades

Recorded 2026-09-01. Extends D81 to the charts.

Choice. CHART_STROKE, CHART_INK and CHART_MOTION in apps/web/src/components/chartTokens.ts hold five line weights, five opacities and three speeds, and the radar, the flag field, the flag bubble and the sparkline read them. A stroke width or a frame opacity written into a chart by hand is drift, and pnpm design:check says so.

The four charts carried nine stroke widths and twelve opacities between them, every one of them chosen in the file that used it. Almost all of them were already within a tenth of each other, which is the tell: nobody was disagreeing, everybody was eyeballing. Collapsed onto five steps each, no weight moved by more than 0.1 and no shade by more than 0.1, and the radar's filled edge, the sparkline's series and the focal bubble's ring now say the same thing at the same weight.

The steps are absolute rather than relative, because every chart here draws a viewBox unit at roughly one displayed pixel. A chart deliberately shrunk below that, like the peer radar in a table cell or the shape inside the field's hover card, shrinks whole and keeps the ratios, which is what it is for.

In-chart text is not in the file. The radar names its axes inside a 260 unit field and the flag field labels its ticks inside a 960 unit one, so one number cannot mean one size across both, and a token that has to be scaled at every use is a literal with a longer name. Those sizes are named inside each chart's own geometry instead, where the field width they are sized against is.

Four things that draw are outside it and say why in their own file: Icon carries Lucide's weight on Lucide's grid, EnvisioningMark is brand geometry, Og draws a 1200 pixel social card in a different medium, and DotField is atmosphere rather than data, which D81 already required it to stay.

Why. A reader learns a picture once. The weight of a line is how a chart says whether something is the frame or the finding, and that only works while every chart on the site agrees. Nothing was holding the agreement together: the weights were close because they were copied, and a copy drifts on the first change nobody compares. The same argument D81 makes for the button and the field, one step down, where it is harder to see going wrong.

The check that guards it warns rather than fails. Every rule in it is a judgment the author is allowed to overrule, a stroke width is not a broken build, and a gate that stops a deploy over a hairline teaches people to route around it. What it buys is that the drift is said out loud on the build where it happened.

Cost. Five steps is fewer than the charts were using, so a chart that wants a weight between two of them has to argue for a sixth step in this file rather than type a number. That is the intended cost and it will be paid by somebody who is right. The shades moved slightly: the median line on the flag field is a little darker, its ticks a little lighter, and the radar's outer ring a little lighter than they were. The warning-only check will be ignored the first time somebody is in a hurry, and nothing will report that it was.

Overturned by. A chart whose natural drawing scale is not one unit to one pixel, which would make an absolute scale wrong rather than merely coarse and would want the weights expressed as a ratio of the field; or evidence that a reader distinguishes fewer than five weights on these surfaces, which would shorten the scale rather than lengthen it.

D104: The drawn network is mounted in this repository, and the ledger stays authoritative

Recorded 2026-09-01. Supersedes the deployment clause of D82. Keeps everything else D82 recorded: the projection, the feed, the drawable rule and the ledger's primacy.

Choice. @envisioning/app is a dependency of apps/web, and the drawn network is mounted at /network, reading the same feed bench institutions writes. /api/institutions/{ISO3} still publishes that feed, now for readers outside this deployment rather than for the drawing itself.

The route is a leaf. The library brings its own header, menu and URL grammar, so /network is reached from a link on each institution page and is not a node in the navigation tree D73 and D80 describe. Its stylesheet is imported in the route's own layout, so it loads there and nowhere else, and theme.css restates NCB's tokens in the shape the library reads rather than sampling its palette.

Why. D82 put the drawing outside the repository because this repository is public and the library is not, so an in-repo dependency makes the viewer uninstallable for anyone without an Envisioning token. That cost was stated and accepted deliberately: one deployment is simpler to keep in step than two, and the project's contributors today are the people who have the token.

Cost. This is the real one, and it is not small.

An outside contributor cannot run pnpm install at all. The failure is not scoped to /network: it is the whole workspace, and the error is a 401 with no explanation of what to do about it. CONTRIBUTING.md describes a build that a reader of the public repository cannot perform.

The dependency adds 246 packages to a repository that previously depended on Next, React and its own core, including d3, zustand, three Radix packages and the AI SDK. /network is a 519 kB first load against roughly 150 kB for every other route.

The peer dependency is an exact pin, react@19.2.4, and this repository resolves 19.2.8. pnpm reports it as an unmet peer on every install. Nothing fails today, and a future release of either side can turn that warning into a break.

Deployment needs the token too. Vercel has no .npmrc for a private scope unless one is provided, so the build fails there until an NPM_RC environment variable carries the registry line and a token that has not expired.

Overturned by. A contributor outside Envisioning who cannot build the project, which is the failure this trades away and the one that should reverse it. Or @envisioning/app published to a public registry, which removes the whole cost. Or the drawing earning its own deployment for another reason, which would return to what D82 described without needing this entry.

D106: The panel is four vendors, activation is explicit, and lost calls are counted

Recorded 2026-08-31. Extends D12, D13 and D14. Governs the first gateway run.

Choice. Three changes to the Delphi run path, taken together because they are the same failure: a panel that looks stronger than it is.

1. Four distinct vendors, one per stance. DEFAULT_MODELS gains mistral/mistral-medium-3.5 beside Anthropic, OpenAI and Google. buildPanel deals models round-robin across four stances, so the previous three-model default gave anthropic/claude-opus-5 both the Institutionalist and the Execution realist seat. Half the panel was one model arguing with itself, and nothing in the run file said so. 2. Activation is always explicit. A run with --max-coverage 1 over the full country set used to set activate true whether or not the flag was passed, replacing latest.json the moment the run finished. docs/PANEL.md and docs/RESEARCH-ROADMAP.md both say to activate after review. The code now agrees with them. 3. Failed calls are counted and published. A provider call that fails after retries is still dropped so one vendor timing out does not cost the other 459 calls. attemptedCalls and failedCalls now travel on the run file, the command warns, and bench validate reports the count and any cell-round that came back with fewer panelists than the panel declares.

The pricing table is also read from the gateway's public model list rather than from vendor pricing pages, so no panel model carries an unverified price.

Why. D12 makes the interquartile range the dissent signal, and dissent above 25 points is how this project finds likely mismeasurement. Every one of the three defects narrows that range without narrowing the disagreement it is supposed to measure. Two stances on one model agree more than two vendors would. A run that lost a vendor to rate limiting returns cells scored by three panelists instead of four, and a shorter list has a smaller IQR. An unreviewed run reaching latest.json publishes both artefacts before anyone can see them. A partial failure therefore reads as consensus, which is the one reading the panel must never produce by accident.

D13 says cost is not the constraint at this scale, so widening the panel to four vendors is not a budget question. The gateway list gives the fourth vendor at $1.50 and $7.50 per million tokens.

Cost. pnpm bench delphi no longer activates anything on its own, so an operator who relied on the old behaviour must add --activate after review. DelphiRunFile gains two optional fields; existing runs parse unchanged and report nothing, which is correct, because a run written before this change has no failure record to publish. Model ids still have to be checked against the gateway before a run: the endpoint is public and needs no key, but a stale id still fails a whole panelist, and that failure is now counted rather than silent.

Overturned by. Evidence that stance and vendor are not separable, which would make the round-robin deal the wrong unit; or a provider contract that makes a dropped call recoverable, which would let the run retry rather than record.

D107: A global body lives once, in a ledger, and a country map reaches it by id

Recorded 2026-09-02. Extends D54, D56 and D82. Answers the question D56 left open about what external means outside Brazil.

Choice. data/institutions/global.json holds the bodies no country owns: the United Nations and its programmes and agencies, the multilateral lenders and the intergovernmental standard setters. Each is one node with the same shape as a country node, at a new level, global, under a new legal nature, international_organization, with an id that starts with global. and a members field listing the registry codes of the benchmarked countries that belong to it. A body with no membership, such as a programme, omits the field.

A country file never holds a global node. It reaches one by id in an edge, which is where a relation between a global body and a national institution belongs, because that relation is a fact about the country with a source of its own: UNDP produces the Atlas of Human Development with Ipea, the IMF's Article IV consultation produces evidence for the Banco Central, IBGE produces Brazil's SDG indicators for the UN statistical system. A relation between two global bodies, such as UNDP's attachment to the UN, lives in the ledger.

attachGlobalInstitutions in packages/core/src/pipeline/institutions.ts is the only place a global node enters a country's network. It attaches every body the country's edges name and every body that lists the country as a member, and it brings a ledger edge along only when both of its ends are attached. The viewer's loader, the explorer feed and the tests all call it, so every surface renders the same network and the file on disk stays a country file. The agenda reads the file alone: its institutionIds point at institutions a country can act through, and a country does not act through the IMF.

Membership renders as a line on the body's profile, in the ledger and in the drawn network's description: how many of the benchmarked countries belong to it, and whether the country being read does. The validator checks that every member is a registry code, that no country file mints a global. id, that a country edge naming a global id resolves against the ledger, and that no country edge joins two global bodies.

Why. D56 recorded that external carries a Brazilian assumption and that the enum survives a second country only because its labels are a lexicon lookup. This is the case it was waiting for. A foreign private university inside São Paulo and the IMF are not the same kind of outsider: one acts inside one country and is a fact about that country's map, the other belongs to no country and would be duplicated 53 times if it were. Duplication is the failure D26 and the registry invariants exist to prevent: 53 copies of the UN with 53 sources and 53 summaries drift, and a reader comparing two countries' maps sees two UNs.

Membership sits on the body rather than as a member_of edge because it is a property of the body, sourced once from its own member list, and because a membership edge would need a national node to hang from. Brazil's map has no node for the state itself; the union is a jurisdiction, not an institution.

Cost. The ledger is a second file every institution surface must read, and a country page now fails when the ledger is malformed, the way it fails when the country file is. The members field is a second kind of fact in a file that otherwise records mechanisms, and it will need refreshing when a country joins or leaves a body. Two of the eight sources, the OECD and IMF member pages, refuse automated fetches, so a live check of the ledger's URLs cannot confirm them.

A body's system and dimensions are a judgment: the WHO under regulation and the UN under strategy_management are defensible and not the only reading. The 11 systems were drawn for a state and a global body fits them loosely.

The drawn network gives global bodies their own jurisdiction cluster, which is not drawable on its own, so they appear only beside the union's network or in a selected institution's neighbourhood.

Overturned by. A body that belongs to no country but acts inside only one, which would make the ledger a place for regional bodies too and need a rule for which file holds them. Or a second country map whose relations to global bodies do not sort into the four families, the same test D56 set. Or readers treating membership counts as a score, which would mean the line should come off the profile.

D108: The registry and its diagnostics are drawn as lanes, and the drawing is a viewer component

Recorded 2026-09-02. Extends D26, D53 and D67. Does not touch D56, D58 or D104: the institution map keeps its ledger, its matrix and its drawn network.

Choice. The viewer publishes one lane field, LaneField in apps/web/src/components/LaneField.tsx, and one page that draws it, /explore in the Method section. A lane is a capability, a dot is an indicator, and the mark says whether the row has data: a filled dot does, a dashed ring is a declared gap and a thin ring is a retired row. A line joins two indicators whose series correlate at REDUNDANCY_THRESHOLD or above. The field has two arrangements. registry packs each lane in registry order, so the length of a lane is the size of the capability's basket. measure moves every dot with data to its absolute correlation with log GDP per capita, draws WEALTH_CORRELATION_THRESHOLD as a rule, marks each lane's own correlation as a tick and parks the unmeasured rows at the left. A dot keeps its node between the two arrangements, so the reader watches it travel. The arrangement is the address: exploreHref and readLaneArrangement in apps/web/src/lib/links.ts are the only place that shape is written or read.

The projection is computed once, by buildIndicatorLanes in packages/core/src/pipeline/lanes.ts, from the registry and the published diagnostics. It reads indicatorVsGdp, wealthAttribution, dimensionVsGdp and redundantIndicatorPairs and nothing else. Nothing in it reaches a score, a confidence or data/out.

The field owns its readout, as the radar and the system matrix do. It always reads one dot and never none, at a fixed height a hover cannot change, and it opens on the dot with the most links because that is the dot the diagnostics have most to say about. The readout names every joined indicator as a control that moves the reading there, so the field is a way into the diagnostics and not only a summary of them.

The encodings are chosen against the charts that already exist. No dot is sized, because size on this site would read as a score. No dot carries a ring for certainty, because that ring belongs to the flag bubble. No dot is a flag, because a flag is a country and these are indicators. The selected dot is the page's one lime mark. Lane names are DOM text beside the field, never SVG text inside it, and the field lays out in pixels from a measured width rather than through a scaled viewBox, so a dot is the same size on a phone as on a desk.

Why. The registry page and the diagnostics page hold the same 69 rows as tables, and neither shows the shape of the whole: that Anticipation and Agency fill their lanes with data while Trust and Shared Purpose are mostly rings, or that every redundant pair the diagnostics find runs through the same five income-tracking indicators. A3 states the second fact in prose. A line between two dots past the threshold states it in one glance, and the measure arrangement puts the whole of the wealth-sensitivity finding on one axis with the threshold drawn through it.

The lane layout was chosen over the force layout the institution network uses because position here means something. A force layout's plane encodes nothing a reader can name, which is what D56 held against the drawn overview. A lane encodes membership and, in the measured arrangement, a number the diagnostics already publish, so the plane carries two facts before any line is drawn.

The projection lives in core rather than in the view so a second drawing can read it. The institution map is the dataset whose shape matches this picture most closely, and a lane version of it is a candidate to replace the closed-source network D104 mounts. That decision is not taken here. The component is built so that taking it later is a data change and not a second component.

Cost. The measure arrangement puts 36 dots on one axis, and the site's rule that every distribution is a FlagField was written for countries on a score axis. This is a different object, indicators on a correlation, and it is drawn with different marks so the two cannot be confused, but it is a second axis chart and a reader who has learned the field will meet a new one.

Six links is a sparse picture. The redundancy threshold is deliberately high, so the drawing says little about pairs at 0.8, and a reader could take the absence of a line for independence. The readout prints the threshold for that reason.

Dots that share a correlation step above or below one another by a fixed offset, so a vertical position inside a lane means nothing and could be read as if it did. The lane is 48 pixels tall to keep that offset small.

Overturned by. A reader taking a dot's position inside a lane, or its absence of a line, for a finding the diagnostics do not make, which would mean the encodings need a legend the field does not carry or the drawing needs to go. Or the lane picture of the institution map being built and the two components diverging, which would mean the projection was not general enough and the second drawing should have been a data change.

D109: The mounted network owns a deterministic, opaque theme boundary

Recorded 2026-09-02. Extends D103 and D104.

Choice. The /network leaf pins @envisioning/app to the dark theme in its own layout. Its route stylesheet maps the NCB palette to the library's surface tokens and repeats the library's --color-* aliases inside the shell. The dense network toolbar's sm:bg-panel/90 and sm:bg-panel/95 surfaces are solid on this route; the graph drawing itself is unchanged.

Why. The library declares its Tailwind aliases at :root as values such as var(--panel). NCB had supplied --panel only on .explorer-shell, so the alias was resolved at the root before the shell could provide its value. The library's nominally solid bg-panel surfaces consequently computed as transparent, and its intentionally translucent toolbar let graph lines bleed through the controls. Event-bff demonstrates the required boundary by pinning its map to a real .light wrapper; the network keeps NCB's dark graph palette but follows the same ownership rule.

Cost. The network has a stable dark presentation rather than inheriting the host OS preference. The adapter depends on the library's current token names and toolbar class names, so a library upgrade needs a surface check. A future visualization can choose a different chrome policy, but it must document that choice at its own route boundary.

Overturned by. @envisioning/app exposing a supported scoped-theme API and an opaque network-toolbar option, which would remove the route adapter; or evidence that the graph is more legible when its controls are translucent, which would justify restoring alpha for this visualization.

D110: The network opens on a national scope, with state institutions one switch away

Recorded 2026-09-02. Extends D82, D104 and D109.

Choice. The mounted network keeps the state layer in the published feed but opens broad views with it hidden. Brazil has 282 state-level institutions and 75 federal institutions, so all levels together make the overview carry more labels than the national structure can support. A route-owned StateLevelToggle writes the library's existing criteria[level.id][$in][] grammar into the address. The dependency adapter applies that criterion to the allInstitutions entity set, so the switch removes state nodes and their lines from the graph rather than merely highlighting them. The generic Level filter remains the full multi-select control for exact scopes.

The default is conditional. A view without a selected institution starts on the non-state levels present in that feed; a view with an institution selected keeps all levels, because a state relation is part of that institution's neighbourhood. A URL-specified level criterion always wins. Other criteria and the dataset selection survive the switch, and the feed, API and source ledger are unchanged.

Why. The network's forcefield visualization reads entities without applying the app's generic filter state. Copying the filter into a second local state would make the modal, the address and the graph disagree. Reusing the app's criteria shape at the route dependency boundary gives the quick switch and the full filter one scope contract, while keeping the source data complete for institution pages and later visualizations.

Cost. The route adapter follows the library's current query grammar and filters the entity collection for every surface mounted under the network, not only the drawing. On a live scope change the library may animate removed labels out before they disappear, so the map settles rather than snapping. The default hides the state layer only for an unselected overview; readers who want the full map must make one explicit switch or clear the level criterion.

Overturned by. @envisioning/app exposing a supported visualization-level applyCriteria or scope API, which would let the route remove its adapter and use the library's own control end to end; or evidence that the 282-state-node overview is a clearer first reading than the national scope.

D111: The network sidebar is the facet navigator

Recorded 2026-09-02. Extends D56, D82, D104 and D110.

Choice. The network sidebar exposes the institution directory plus one destination for each reusable facet in the feed: systems as clusters, levels as tags and jurisdictions as clusters. Each facet has an index and a detail page whose related-institutions list is joined from the same entity records. Facet cards show their related-institution count. The institution list keeps the exact multi-select filters and its Group by control, and the forcefield keeps its Group by control for changing the spatial arrangement. The sidebar therefore browses a facet, while the filters narrow a view and grouping changes how the current set is read.

Why. Event-bff's Envisioning implementation treats clusters and tags as first-class navigation surfaces instead of hiding every useful dimension in a filter modal. NCB's feed already declared these relationships and the graph already knew how to cluster by them, but only systems were reachable from the sidebar. Giving the three facets stable entry points makes the same data legible as a directory, a tag index or a jurisdiction map without duplicating the ledger or inventing a second filtering model.

Cost. The sidebar grows from three to five destinations and is therefore more dependent on the library's compact-menu and navigation-map behavior. A facet page is a browsing surface, not an independent dataset: its counts and related institutions change with the published feed, and it does not replace the exact filter grammar. The network's facet pages also inherit the mounted library's route and token contracts.

Overturned by. Evidence that readers cannot distinguish browsing a facet from filtering the institution set, or that the sidebar becomes unusable at the supported breakpoints; or a supported Envisioning facet-navigation API that provides the same destinations without route configuration.

D112: The hero's dot motif breathes, behind the reader's motion setting

Recorded 2026-09-02. Supersedes the still-only clause of D81. Keeps every other clause of D81: the alphas, the geometry, the one dark band, and the rule that the motif is atmosphere and never data.

Choice. DotField draws the parent brand's breathing motion. The bubble around each centre grows and shrinks on a sine along the same diagonal phase the still frame already froze, on the parent's 3200ms cycle, at the alphas D81 set. The motion is gated three ways. A reader whose system asks for reduced motion gets the still SVG pattern and no script runs for it. The server renders that same still frame, so it is what every reader sees first. The canvas takes over only when motion is allowed, and it stops drawing when the band leaves the viewport or the tab is hidden. Its first frame is the still frame, because both read one phaseAt and time starts at mount, so the handover does not jump.

Why. D81 recorded that a research surface does not need a background that moves and that a still frame costs no JavaScript. Both are still true and neither was the reason a reader asked whether the motif was animated. The still frame is one moment of a motion the parent site runs everywhere, and a reader who knows the parent reads it as animation that has stopped, which is a broken thing rather than a quiet one. Drawing the motion restores the motif to what it is, and gating it behind the reader's own setting keeps the quiet version for the readers who asked for it.

The alphas do not move. At 0.05 and 0.22 the breath is felt in peripheral vision and is hard to see when looked at, which is the test D81 set for the still frame and the test the moving one has to pass too.

Cost. The hero now runs a requestAnimationFrame loop while it is on screen, on the one route that opens the site. The loop draws about 300 circles a frame at desktop width and stops off screen, so the cost is a few milliseconds a frame for as long as the band is visible and nothing after. The still pattern is kept for the reduced-motion reader, so there are two drawings of one motif in one file, and a change to the geometry has to reach both, which phaseAt and bubbleAt are there to make one change.

Overturned by. A reader finding the motion distracting at these alphas, which would mean atmosphere has become texture and the still clause of D81 should return. Or a measured cost on the front page's interaction readiness that the loop causes, which would mean the canvas has to wait for idle before it starts.

D113: The front page shows a surface rather than naming it

Recorded 2026-09-02. Amends the one-link clause of D75. Extends D53, D58 and D108.

Choice. Where a section of the site is a drawing, the front page draws it and the drawing is the link. Two modules changed. The wealth module, which states how far each capability tracks income, now carries the lane field in its measured arrangement as a still picture, and the picture opens /explore. A new module carries Brazil's system matrix as a still picture, cells shaded by count and nothing else, and the picture opens /network. Both pictures are the same components the pages draw, in a non-interactive mode: no readout, no hover, no keyboard, one link around the whole. LaneField gained interactive={false} for it, the way Radar has inside a grid card, and MatrixPicture in apps/web/src/components/MatrixPicture.tsx is the matrix without its readout, reading the same MATRIX_FILL the institution page reads.

Why. A sentence that names a page tells a reader it exists. The page's own picture, at the size the front page has room for, tells the reader what it is, and that is the difference between a visit and a link that was never followed. D75 wrote each module as a sentence, a live number and one link out, and a link that is a picture keeps the count while carrying more.

Brazil is drawn because it is the one country with a map. The front page does not choose it; it has nothing else to draw. When a second map lands, this module becomes a choice and needs a rule, which is why the module is gated on the file and not on a list.

Cost. The front page now loads the Brazilian institution file and the global ledger on every request, beside the index, the diagnostics and the evidence corpus. It is a few hundred kilobytes of JSON read from disk, and nothing is cached, which is the cost every other page pays too.

The lane field is a client component, so the front page ships it whether or not the reader points at it. It is drawn without a readout, which makes the still picture the picture's whole meaning: a reader who cannot point at a dot reads nothing about any one indicator here and has to open the page.

The matrix picture carries no axis names. Ten unlabeled rows against ten unlabeled columns is a texture until the reader opens the page, and the section hint has to carry what the axes are.

Overturned by. A second institution map, which makes the Brazilian module a choice the front page has no rule for. Or a measured cost on the front page's response time from the added reads, which would mean the matrix should be computed once by bench institutions and read from data/out.

D114: A count is drawn as a count, and a year as a position in time

Recorded 2026-09-02. Extends D103. Serves D45 and D100.

Choice. Two marks, both second encodings of numbers the viewer already printed. CoverageMark draws one tick per indicator row a capability is measured against, filled where the row was observed for this country, with a notch at the coverage floor. RecencyTick draws a rail from the first year the ingest fetches to the current one, with a tick at the year a value was observed and a quiet line trailing it for how long ago that was. Both print the number beside the mark and neither replaces it.

Coverage reached the reader as 0.71. Seven rows cannot produce a hundred values, so two decimals claimed a resolution the count does not have, and the ratio hid the thing a reader most needs from it: how far a capability is from the two rows a score requires. Below the floor it was worse. The page printed the words "not measured" and the count that explains them existed only in screen reader text, so the reader who could see the page was told least.

Recency is one of the three factors behind every confidence and it reached the reader as a decimal in a sortable column, or as a bare year on an indicator record. Both are unreadable without a scale: 2011 means nothing until it sits next to what a current round looks like. The rail is that scale, and the length of the trailing line is the staleness the decimal was trying to say.

Both denominators come from countedForCoverage in the registry, which the scorer also uses, so the mark and the confidence figure beside it can never be counted against different denominators. A retired row is in neither.

Why. These are the two things this dataset is least certain about, and both were rendered as the kind of number a reader skims past. A count drawn as a count is read without being parsed: five ticks lit of seven is a glance, 0.71 is arithmetic. The floor notch is the same argument for D45, which withholds a score below two observed rows and until now gave the reader no way to see how close a capability was to publishing one.

Cost. Two more marks to learn, on tables that already carry a score band, a confidence chip and a trend arrow. The coverage mark stops being countable somewhere above a dozen rows, so a capability that grows past that needs a different form rather than a tighter pitch. The recency rail's right edge is the current year, which means every mark on the site shifts by a hair each January, and a screenshot taken now will not line up with one taken next year.

Two places that were named as candidates and refused. /gaps counts registry rows with no dataset, which is a different claim from rows observed for a country, and one mark meaning two things is worse than two marks. The indicator peek panel already draws a field above its list and a bar in every row, so a third mark there would crowd a reading rather than open it.

Overturned by. A capability with enough rows that the ticks stop being countable, which would want a different form for coverage and not a smaller pitch; or evidence that readers take the trailing line for a bar rather than for elapsed time, which would make the rail worse than the year it replaced.

D115: Out of frame is read against the values that built the frame

Decision. buildFrame keeps the unclipped extremes of the values that set it, as observedMin and observedMax, and scoreAgainstFrame marks a value out of frame when it lies beyond them, as well as when its clipped position falls outside the endpoints. Dataset 6.1.1.

Why. D47 promises that a current value can never fall outside a frame its own country helped build, and that a historical value beyond it clamps and is flagged. The flag was read from the clipped value only. When a current outlier is winsorized onto the upper fence, that fence is also the top endpoint, so a historical value further out was clipped to exactly the endpoint and never reported. The synthetic test in normalize.test.ts found it. A value between the fence and the most extreme current value is winsorized exactly as that current value is, and stays in frame: the frame was built with a country there.

Cost. Four more clamped cells across the momentum baskets in diagnostics.json. No score and no current cell moves, because a current value sits inside the observed extremes by construction. The two new fields live on the internal Frame type and are not published.

Overturned by. A reason to treat the fence rather than the observed extreme as the edge of the frame, which would mean flagging every winsorized historical value, and the current outlier's own winsorized value with it, breaking the D47 promise from the other side.

D116: The sources page prints the check series too

Decision. /sources lists every behavioural check from checks.ts in its own section, with the series code, its database and the full request built by the same worldBankSeriesUrl call the indicator example uses. This closes the cost D60 recorded, that the fetch the page prints back was short by the check calls.

Why. The page's claim is that a reader can repeat every call the benchmark makes. A check is fetched on every ingest, so leaving it off made that claim false by one series.

Cost. A second table on a page that already carries the indicator one, for series that never enter a score. The section says so in its heading and links the glossary entry rather than defining a check again.

Overturned by. A check fetched from somewhere other than the World Bank, which the request builder cannot print, and which would need the page to print another source's call shape before it could stay complete.

D117: Research is chosen by two objectives and triaged before values are fetched

Decision. Every research task names the objective it moves and reports the move in its handoff. O1, informative: every dimension reaches a mean confidence of at least 0.40. O2, separable from wealth: no dimension sits above r = 0.70 against log GDP per capita. A guardrail sits beside them: mean confidence must not come to track wealth. docs/RESEARCH-ROADMAP.md holds the queue in that order and a desk triage step that every candidate passes before any value is fetched: the publisher's own coverage ceiling against the 53, whether the values will spread, whether the series is a stock that money buys, and whether one adapter or one file serves the whole frame. A candidate that fails triage is recorded in one paragraph and stopped. The queue opens with sweeps of the two source families that already have adapters and full coverage, V-Dem and the Joint EVS/WVS release, ahead of single indicators.

Why. The claim under test is that capability is separate from wealth (docs/WHY.md, D1). At dataset 6.1.2 four dimensions miss O1 (Trust 0.21, Experimentation 0.23, Shared purpose 0.26, Coordination 0.36) and four miss O2 (Anticipation 0.87, Agency 0.85, Adaptability 0.82, Learning 0.73), and the rows carrying O2 are mostly class I diffusion stocks. The queue was ordered by gap instead, and TRUST-2 showed the cost: a full value preflight and a search of every national judiciary ended at 13 of 53, a result the desk ceiling of 26 already implied, on a ratio that sits between 0.87 and 1.02 in 12 of the 13 countries it reached. Harmonised administrative statistics across this frame exist mostly where a regional body pays for them, so indicator-by-indicator search drifts toward rich-country sources, which is what the guardrail watches. Mean confidence against log GDP is r = 0.34 across 51 countries today.

Cost. The two targets are judgment, not derived: 0.40 separates the four thin dimensions (0.21 to 0.36) from the rest (0.47 and up), and 0.70 reuses the probe's single-series screen at the dimension level. Candidates that would raise confidence only for rich countries, PISA among them, are parked, so Learning's O1 standing gains nothing from them for now. A triage paragraph can kill a series that a full preflight would have rescued.

Overturned by. A dimension that reaches both targets and still fails the claim, for example by pairing with another dimension in duplicateDimensionCandidates in diagnostics.json, which would show the objectives measure the wrong thing; or a triaged-out candidate that a later preflight shows would have cleared the half-frame screen, which would show the triage questions are too strict.

D118: A row is chosen for what it measures, and its income correlation is reported, not used to choose it

Recorded 2026-10-01. Supersedes the O2 target in D117, the wealth condition in D52's probe pass and the acceptance rule in D42. Keeps D42's diagnostic.

Decision. A candidate is accepted, held as a check, left as a gap or retired on its construct: what it observes, whether that is a capability or a stock that money buys, and whether it is behaviour, an outcome or a perception. The source memo states that argument, with its date, before any value is fetched. The correlation with log GDP per capita, the wealthAttribution delta and the dimension's correlation after the change are still computed, and every decision entry and handoff prints them. They are findings, and they no longer pass or fail a row.

In practice:

• pnpm bench probe reports a series that tracks log GDP at 0.70 or more as a flag beside the verdict, not as a failure. Coverage, recency and spread still fail a series.
• O2 becomes a reported outcome. Each release prints every dimension's correlation with income, and the research queue no longer aims at a threshold. O1, the 0.40 confidence target, and the guardrail stay: neither chooses rows by what they say about income.
• The roadmap's wealth-link work becomes a construct audit. A diffusion stock such as secure servers or broadband subscriptions is retired if it does not observe the capability its dimension names, and kept if it does, whatever its correlation.
• Exclusions that rested on income alone are reopened for a construct review. The bribery_incidence check (D60) is the first: it records experience, not reputation, and D60 held it out only for its wealth contribution. D44's retirement of homicide_rate gave construct reasons too and stands.
• The next real use of the benchmark is a report on Brazil's adaptability, so the queue serves Adaptability first. The frame does not change: every source is still tested against all 53 countries, and Brazil has no special treatment in the code.

Why. docs/WHY.md calls the benchmark a test of whether capability is separate from wealth, and names the failure: the nine dimensions collapse into one factor that tracks GDP per head. If rows are admitted or dropped by what they do to that correlation, the test cannot fail, because any row that would make it fail is removed. D42 and D44 show the mechanism. IC.FRM.CORR.ZS was wired and reverted in one session because it moved Trust from 0.385 to 0.619, and D60 held out a behavioural bribery measure for the same reason. Both were reasonable steps under the rule as written. Repeated across the registry, though, the rule makes a low income correlation a product of the selection. Choosing by construct and letting the correlation land where it does turns the dimension correlations back into evidence. The wealth residual (D68) still shows what a dimension carries beyond income, and it does so honestly only if the rows were not chosen to make it large.

Cost. Some dimension correlations will rise, and a dimension may cross 0.70 and stay there. That is published as a finding against the claim. A construct argument is softer than a number and easier to dispute, so the memo states it before the values are seen. The rows chosen under the old screen are not re-admitted wholesale: each reopening is its own reviewed change, and that is slow. D117's triage keeps its ceiling, spread and cost questions, and its wealth question becomes the construct question above.

Overturned by. An audit of rows decided after this entry showing that the construct arguments, written before the values, still drift toward rows that lower the income correlation. That would show the pre-commitment is not working. It would also be overturned by rows admitted under this rule that reviewers agree are levels of spending or adoption under another name, which would show the construct question is too loose to stand in for the number.

D119: Export concentration from UNCTADstat fills export diversification

Decision. export_diversification moves from gap to adapter. The value is UNCTADstat's Concentration Index (US.ConcentDiversIndices, flow 02, exports): a normalised Herfindahl-Hirschman index of the merchandise export basket across SITC Rev.3 3-digit products, 0 when exports are spread evenly and 1 when one product is everything. It is stored as published, so the row becomes direction: 'lower_better' with the unit "index 0-1, lower = more diversified", rather than being inverted in the adapter. The adapter fetches the keyless bulk 7z, extracts its one CSV with bsdtar, joins UN M49 codes to the registry, and emits the pinned year 2025 only. The file is pinned by the SHA-256 of the CSV (release stamped 2026-06-24), so a publisher refresh fails the fetch until someone re-pins it. UNCTAD's footnote travels in each observation note: 15 of the 53 values are marked Estimated (mirror data from partners). The Diversification Index in the same file is not used: it measures distance from the world basket, which is a resemblance, not a spread.

Why. The row is accepted on its construct, decided before its wealth correlation was read. Adaptability asks whether a country can absorb a shock and reallocate. A basket spread across many products is the standing result of past reallocation and the exposure a single price or demand shock meets, which is the construct the registry already declared for this row. It is class C, an observed outcome of the economy's structure, and not a diffusion stock that money buys directly. The source covers all 53 countries at 2025 from one international publisher, so it adds evidence without favouring rich countries.

Reported as findings, not tests. The row's own correlation with log GDP per capita is r = 0.445 (n 51; the stored HHI is direction-adjusted so a higher normalised score is more diversified). Its wealth-attribution delta is +0.021: with the row Adaptability sits at r = 0.839 (Spearman 0.837, n 51), without it 0.818, so the row does not move Adaptability toward O2 and slightly away from it. It forms no redundant pair. Adaptability's mean confidence rises from 0.469 to 0.588, observed rows from 4 to 5 for every country, and mean confidence against log GDP across all dimensions stays at r = 0.34, so the guardrail holds.

Cost. The index reads a product mix, not the capacity to switch. Switzerland (0.364, gold and pharmaceuticals), Ireland (0.330, pharmaceuticals) and Singapore (0.271, re-exports) score as concentrated because a few lines are worth a lot, not because they are fragile. Commodity exporters such as Venezuela (0.762) and Nigeria (0.619) move with prices from year to year even when the basket does not. It is merchandise only, so services-led exporters read narrower than they are, and 3-digit SITC hides diversity inside a product line. The registry note says all of this. The history from 1995 is in the file and is not emitted, so the row has no trend and no discrimination trend yet. The fetch depends on bsdtar being on the path. The data is CC BY 3.0 IGO and must be cited as the UNCTAD Data Hub.

Overturned by. A source that observes reallocation itself, such as entry into new export products or the speed a basket recovers after a price shock, at comparable coverage, which would answer the construct more directly and replace this row; or evidence that the high-value-line artefact moves more countries than the three named, which would make the row read wealth structure rather than exposure and return it to gap.

D120: Long-term unemployment is wired from ILOSTAT behind a plausibility gate

Decision. long_term_unemployment_share (Adaptability) moves from gap to adapter. The adapter reads ILOSTAT's DF_UNE_TUNE_SEX_AGE_DUR_NB in one SDMX call and derives the share as unemployed 12 months or more over the unemployed with a stated duration, both sexes, age 15 and over, a labour force survey preferred where ILOSTAT holds more than one. Lower is better, raw percentage. Before emitting, it runs a plausibility gate on every country, naming none: it drops a country-year under 3%; every year of a survey whose median for that country is under 3%; a run of one or two observations that sits more than 15 points beyond both neighbours when those neighbours agree within 15 points; and a latest value more than 15 points from the one before it, which waits for the next year to confirm it. A jump the next year keeps is a level shift and stays. It emits the latest surviving year and logs every dropped value with its country, year, value and reason. The run of 2026-10-01 emits 44 of 53 countries, holds KOR, MEX, PER, PHL, SLV and URY with no surviving year, and drops 78 country-years; Brazil is 30.2% in 2025 from PNAD Contínua. The list is in docs/research/adaptability/ILOSTAT-LONG-TERM-UNEMPLOYMENT.md.

Why. Construct first. The row observes whether people who lose work find new work, which is reallocation, the centre of Adaptability, and it is behaviour rather than a stock money buys. That is the reason it is accepted. Its two readings are recorded in the registry note: a high share is slow reallocation where the unemployment rate is also high (South Africa, Kenya, Nigeria) and a small residual pool where the rate is low (Switzerland, Japan), so it is read beside unemployment_rate. Some questionnaires cannot record a long search, and their published shares (under 1% in every year for Korea, the Philippines and Peru, and for Uruguay outside 2021 and 2022) describe the instrument. A rule applied to every country keeps the exclusion auditable and stops it from becoming a list of countries the benchmark disliked. The survey-median clause exists because a year-by-year floor alone let one year of an otherwise sub-3% questionnaire through (Mexico 2022 at 3.1, El Salvador 2021 at 4.6), and those would have taken the best cells in the frame. As a reported finding, not a test: the row's normalised score correlates with log GDP per capita at r = 0.355 (raw share -0.355, n 42), its wealth-attribution delta is 0.005, and on the local run Adaptability moves from r = 0.818 to 0.824 (n 51) while its mean confidence rises from 0.469 to 0.558.

Cost. The thresholds, 3% and 15 points, are judgment. A real one-year excursion of more than 15 points is dropped as a spike, and a real level shift in the latest year is held back for one release, emitting the year before. The survey-median clause goes beyond the owner's two-clause rule and removes values that clear 3%. The gate cannot see a break with no spike: Ethiopia is emitted at 53.9 for 2013 because its only other year, 2021, falls under the floor. ILO's own unreliable flag is recorded in each note and not used. Four emitted countries come from household surveys rather than labour force surveys, Brazil among them, and Argentina's survey is urban only. ILOSTAT revises in place, so the retrieval date is the release identifier.

Overturned by. A national statistics office or the ILO documenting that a held series is a valid measure of long-term unemployment, which would make the floor a rule that removes real values; a second source for the same countries, such as OECD or Eurostat, that disagrees with the gate's survivors by more than the gate's own 15-point tolerance; or a review of unemployment_rate beside this row showing the two readings cannot be separated in practice, which would make the row a check under D60 instead of an indicator.

D121: V-Dem polarization is published as a behavioural check, not scored

Recorded 2026-10-01. Extends D60 and D83. Supersedes the overturn clause of D116. Issue #22.

Decision. V-Dem's political polarization item is published beside Shared Purpose as a behavioural check under D60 and enters no score. The check is political_polarization in checks.ts, read from v2cacamps_osp, the measurement-model estimate on the codebook's 0 to 4 scale, where 0 means supporters of opposing camps generally meet in a friendly manner and 4 in a hostile one, direction lower_better, V-Dem v15 (2025-03-04), year 2024, tier expert_panel, CC BY-SA 4.0. It covers 53 of 53 countries.

The registry gap political_polarization in indicators.ts stays a gap. The measurement Shared Purpose wants, hostility between camps that are free to exist, is still unmade, and this item is not it. The check shares the gap's id on purpose, and its observations sit under __check__political_polarization, so the two never meet in a frame. The branch polarization-vdem (commit 1c0f6b5) that scored the item is not merged; its adapter work is.

The item is not in the Core archive D83 pinned, so the V-Dem adapter now pins the Full+Others archive of the same release, V-Dem-CY-FullOthers-v15_csv.zip, streams its 400 MB CSV line by line and reads a table of variables, one per observation id. Its v2x_cspart values match Core's for all 53 countries, so civil-society strength does not restate. The observation file keeps its name, vdem-cy-core.json.

Checks are no longer World Bank only. CheckDef gains ingest, worldbank by default or adapter, and pinned, which names the dataset, the archive URL, the file inside it, the column and the year. An adapter check must carry pinned, and checks.ts refuses one that does not. The World Bank ingest and worldBankCheckSeries skip adapter checks, the adapter emits them under CHECK_PREFIX, and the scorer, behaviouralChecks, the report and the capability and country pages read them unchanged. /sources prints the World Bank request for a World Bank check and the pinned archive URL, file, column and year for an adapter check.

Why. The project judges a row by what it measures first and reports its income correlation beside it; the correlation is evidence, not the gate. On construct this item fails in a way no numeric screen sees. Scored on the branch, it passed every gate: 0.335 against log GDP per capita, 0.565 at most against any scored row, 53 of 53 countries, and it lifted Shared Purpose from 47 to 52 published countries. But a low reading has two causes the number cannot separate. On V-Dem's own regime classification the 2024 values form a U: liberal democracies average 1.77 and closed autocracies 1.85, while electoral democracies and electoral autocracies average 2.80 and 2.98. Scored, the item raised the United Arab Emirates by 11.1 points and Rwanda by 11.5, and published Singapore and Vietnam on the strength of their calm. Low measured polarization under repression is not people seeing themselves in a common project. It is A5 inverted: there a perception composite penalised political uniformity, here an expert item rewards it, and the spec refuses both readings. A13 records it.

The item is still worth showing. It is current, full-frame, inspectable and asks the right question of every country where camps may compete. A check is the D60 shape for a series that is real and disqualified, and the reason travels with the number.

Cost. Shared Purpose stays at two rows, 47 published countries and mean confidence 0.260; the coverage the item would have bought is declined. No score moves: all 477 country-dimension cells are identical with the check present. D60 framed a check as a series kept out on income. This one is kept out on construct, and its wealth correlation, -0.335 on the published value in behaviouralChecks, is reported and is not the reason; the glossary entry and the diagnostics comment now say a check can fail either way. A published number outside the score will be quoted as a finding for the closed regimes it flatters, and the attached note is the only guard. The adapter downloads 26 MB instead of 15 MB and depends on the Full+Others member name as well as on unzip. The dataset version does not move: no scored row, field or country changes, and checks is an existing field.

D116 amended. D116's overturn clause named a check fetched from somewhere other than the World Bank, which the request builder could not print. This decision is that check. D116 stands as amended: /sources stays complete by printing each check's own call shape, the World Bank request where ingest is worldbank and the pinned file, column and year where it is adapter. D116's clause is superseded by this decision's.

Overturned by. A V-Dem reading conditioned on competition existing at all, or a behavioural Shared Purpose row (civic participation, volunteering, voter turnout) that agrees with this item outside the closed regimes, either of which would let the item or a variant score and move it to indicators.ts; a V-Dem release that changes the question or drops coverage below half the frame; or evidence that readers take the check for a score, which under D60 would retire it. For /sources: a check whose source has neither an API request nor a pinned file to print, which would leave the page short of a call again.

D122: What a country has is published beside its score as a condition, not scored

Recorded 2026-10-01. Extends D118 and D60. Dataset 7.0.0. Memo: docs/research/CONDITIONS-AUDIT.md.

Decision. The registry gains a second role. A row with role: 'capability' observes the country doing what its dimension names, a behaviour, a throughput, an outcome or performance relative to resources, and it is scored when it has data. A row with role: 'condition' records what the country has to work with, a stock of infrastructure, access, money, people or enrolment, or income itself. A condition stays in indicators.ts and keeps its ingest route, so it is fetched on every World Bank ingest. It is not scored: isScored is false for it, so it leaves the frame, the dimension mean, the coverage denominator, the confidence, the momentum basket, the redundancy test and wealthAttribution, exactly as a retired row leaves the denominator under D100. indicatorsFor no longer returns it and conditionsFor does.

It is published beside its dimension as conditions on every DimensionResult: the value as the publisher wrote it, unit, year, source, and its rank among the n countries that have it, read in the row's direction on the registry transform. No condition is put on the 0 to 100 scale, for the reason D60 gives for checks: a number on that scale reads as a score. indicators/{id}.json is still written for each condition, with role: 'condition', no normalised value and countries in rank order. diagnostics.conditions correlates every condition with log GDP per capita and with its own dimension's published score. The country page, the capability page, the Brazil layer, the agenda in both lexicons, /sources, /indicators and /explore show conditions as a separate layer. On /explore a condition is a square in its capability's lane, placed by its correlation with income in the measure arrangement, never flagged as a wealth proxy and never joined by a redundancy line, because both are verdicts on scored rows.

Tier A of the audit moves, 10 rows: rd_expenditure_gdp, researchers_per_million and secure_internet_servers beside Anticipation; internet_users, account_ownership and domestic_credit_private beside Agency; tertiary_enrollment and education_expenditure_gdp beside Learning; broadband_subscriptions beside Adaptability; labour_productivity beside Building. Each row's note says why, on construct. Tier B (business start days and procedures, vocational share, labour force participation, transmission losses, income inequality) stays scored, and so do the two borderline rows, sci_articles_per_million and human_capital_index. A Tier B row moves only on its own decision entry, and only once its dimension has a capability row to take its place: as a block Tier B leaves Agency and Shared Purpose below the D45 floor in every country.

Why. docs/WHY.md says the benchmark tests whether capability is separate from wealth. D118 chose rows by what they measure, and the construct audit then found that ten scored rows measure what money buys rather than what a country does with it. Averaging them into a capability score answers the question the benchmark refuses, how rich is this country, under the name of another. Retiring them would have thrown away the other half of the reading. The Brazil adaptability report needs both sides: what a country has to work with, and whether it turns that into capability. The second correlation in diagnostics.conditions is that map. It is the reason the layer exists.

Measured on the 7.0.0 run against 6.2.0, same observations, no re-ingest:

Coordination, Trust, Experimentation and Shared Purpose do not move. Across the eight dimensions with a complete panel (46 countries, Trust left out for coverage), the first factor's share of variance falls from 0.62 to 0.50 and the mean absolute correlation between dimensions from 0.55 to 0.42. On all nine (33 countries) the share falls from 0.60 to 0.50. The guardrail, mean confidence across dimensions against log GDP per capita, falls from 0.39 to 0.32 (n 51). Redundant indicator pairs fall from six to none. Cuba loses its Anticipation and Learning scores, each now on one observed row. 261 of 477 country-dimension cells change.

The conditions themselves, r against log GDP per capita and against their dimension's score:

Most conditions go with income more closely than with the capability they sit beside, which is what the split was for. Anticipation is the exception: its conditions go with its score as closely as with income, and its two capability rows track income at 0.87 by themselves. That is published as a finding against the claim, not corrected.

Cost. Confidence falls where conditions carried it. Agency (0.38) and Learning (0.37) drop below the O1 target of 0.40, and Anticipation scores on two rows. That evidence was always this thin; the stocks hid it. Agency now rests on new business density and two Doing Business series frozen at 2019 (A6), and its scores move most: Nicaragua rises 31.8 points, Rwanda 22.4, Japan falls 20.2 and Venezuela reaches 0. The GDP-stripped test now empties Anticipation, because both of its remaining rows sit past the wealth threshold. Published scores move in five dimensions and DimensionResult gains a field, so this is a major version under D37 and 6.x numbers are not comparable. A condition beside a score will be read as part of it by some readers, as a check can be; the panel says "not scored" in its label and the note travels with the value. The rule is a construct judgment, and the audit was written with the income correlations in view, which the memo discloses. A row's role can be argued, and each argument costs a decision entry.

Overturned by. Evidence that a condition observes the capability after all, for instance a use measure that tracks the stock so closely that the two cannot be separated, which would return that row to the score by its own decision. Readers or consumers treating the conditions layer as a second score, which would mean the layer does the harm D60 warns of and the rows should move to the indicator rows with status, or out of the published dimension altogether. Or a capability row for Anticipation, Agency or Learning whose arrival shows that the confidence lost here was a cost of the split rather than of thin evidence, which would reopen whether the move should have waited for replacements, as Tier B does.

D123: Bribery incidence is scored in Trust, because it observes whether the rules hold

Recorded 2026-10-01. Reopens the check D60 created, under D118. Supersedes D60's reason for holding it out; D60's mechanism for checks stands.

Decision. bribery_incidence, World Bank IC.FRM.BRIB.ZS, moves from checks.ts to the indicator registry as a scored row in Trust's institutional family, lower is better. It is the share of firms asked for at least one bribe across six public transactions: utilities, permits, licences and taxes.

Why. Trust asks how much cooperation is possible beyond immediate personal networks, and its high end reads "strangers cooperate on the strength of the rules". This series observes whether the rules hold where a firm meets the state as a stranger. It records experience, not reputation: the respondent is asked whether it was itself asked, which is what separated it from the perception composites D23 retired. D60 judged it valid on that ground and held it out only because it carried income. D118 removes that reason. Its income correlation (about 0.66 alone when D60 measured it) is now published as a finding.

It covers 50 of 53 countries, all but Argentina (2017), Nicaragua and Honduras (2016), South Africa (2020) and Venezuela (2010) at 2023 or later, because the Enterprise Surveys now run in high-income economies too. It gives Trust's institutional family a second observed row beside contract enforcement, which D57 asks for before Trust can be read.

Cost. Respondent reticence. Firms in some systems do not report a bribe request to a survey, so a low value can mean clean transactions or a cautious answer. China reads 0.14% and Korea 0.02% while Vietnam, surveyed by the same programme, reads 31%. Where it bites, the pattern is the A13 trap in survey form, and it is recorded as an artefact rather than corrected, because the Enterprise Surveys' own reticence adjustment is not published per country. Venezuela's 2010 value is old and the recency term marks it down. Trust's correlation with income rises; the release prints by how much.

Overturned by. Evidence that reticence, not experience, drives the cross-country ordering: for example the Enterprise Surveys' published reticence indicators placing the lowest-reporting countries among the most reticent. That would make the row a perception of risk rather than an observation, and it would return to a check with that reason.

D124: Research citation impact is wired from OpenAlex as a top 10% share

Recorded 2026-10-01. Issue #23. D122 and D123 are being written on other branches; this entry takes the next free number after them.

Decision. research_citation_impact (Learning) moves from gap to adapter. The value is the share of a country's articles and reviews, published 2019 to 2021 and stamped 2021, that OpenAlex places in the top 10% most cited for their subfield and publication year (citation_normalized_percentile.is_in_top_10_percent), divided by the same share across every work with an institution country. A work counts for every country any author's institution sits in (whole counting). Unit "ratio to world average", higher_better, class O, tier academic_survey. The definition is reworded from "field-normalised citation impact" to say this.

The adapter openalex-top10-share-v1 makes two grouped calls by authorships.institutions.country_code and two baseline counts, with corpus=core written into each request. It asserts that all 53 countries appear in both grouped results and counts any that does not with two per-country calls; on 2026-10-01 none was missing. OpenAlex has no version parameter, so the observation file carries the pin under openalex: every request, the retrieval date, the totals and each country's two counts. The values derive from those counts, a rescore never touches the network, and a refetch is the explicit pnpm bench openalex fetch, diffed into revisions.json. The mailto and any API key are sent and never stored. Brazil is 0.602 (62,795 of 606,599 works, 10.35% against 17.20%). The memo is docs/research/learning/OPENALEX-CITATION-IMPACT.md.

Why. Construct first. Learning asks whether a country absorbs and produces knowledge. Whether the research it produces is used by others, relative to what is normal in that field, is an outcome of that, not a stock of spending, enrolment or researchers that money buys directly. A share is size independent, the field and year normalisation is OpenAlex's own, and a count of flags cannot be moved by one paper the way a mean FWCI can. The ratio to the affiliated world corrects a level OpenAlex sets against a pool dominated by unaffiliated, uncited works (affiliated works sit at 17%, not 10%); within one window it rescales every country alike and changes no score. Dataset 7.0.0 (D122) moves Learning's stock rows to the conditions layer and leaves Learning thin, and this row measures the dimension rather than its inputs.

Reported as findings, not tests, from a local run on dataset 6.2.0: the row's normalised score correlates with log GDP per capita at r = 0.588 (n 51), its wealth-attribution delta is 0.063, and Learning moves from r = 0.731 to 0.794 (Spearman 0.733 to 0.813) while its mean confidence rises from 0.495 to 0.586. It forms no redundant pair; the closest rows are interpersonal_trust at 0.799 and sci_articles_per_million at 0.796. The prior wealthProxyPrior moves from 0.3 to 0.6, which is what the issue memo measured.

Cost. Whole counting lifts small countries whose researchers co-author with larger systems: counting only single-country works, Panama falls from 21.25% to 3.14%, Kenya from 19.83% to 5.94% and Estonia from 32.89% to 18.26%. The domestic-only share is not published; it is the obvious behavioural check under D60 and is left as follow-up. 37% of articles and reviews with a percentile in the window carry no institution country, and missing affiliations thin small producers most; Haiti rests on 389 works and no floor is applied. Conference papers are out, which understates computing-heavy systems. The value drifts: OpenAlex recomputes citations, percentiles and affiliations continuously, and on the day of the fetch some counts moved within two minutes. A later fetch will restate every value, which revisions.json will show, and the number published is the number read on the stamped retrieval date. The data is CC0; the fetch costs four credits of the keyless $0.10 daily budget.

Overturned by. A fractional-count or domestic-only reading, from the OpenAlex snapshot or the API, that reorders the frame enough to show whole counting is reading collaboration networks rather than use of a country's own research, which would swap the construct or make this row a check; affiliation coverage shown to bias the share by country income or region, which would make the row read indexing rather than research; or OpenAlex changing the percentile's pool or definition, which would need a new adapter version and a new entry.

D127: Perceived control is scored in Agency from the Joint EVS/WVS A173 mean

Recorded 2026-10-01. Extends D64's adapter under D117 and D118. D125 and D126 are being written on other branches; this entry and D128 take the next numbers after them.

Decision. perceived_control (Agency) moves from gap to adapter. The value is the publisher-weighted mean, on the 1 to 10 scale, of A173, "how much freedom of choice and control you feel you have over the way your life turns out", from the same pinned Joint EVS/WVS 2017-2022 v5.0.0 results PDF as interpersonal_trust. Class P, higher_better, tier academic_survey, unit mean 1-10, publisher Joint EVS/WVS. The adapter now reads an item table, JOINT_EVS_WVS_ITEMS, the way vdem.ts reads variables; country mapping and D64's hold rule are shared, so Germany, the United Kingdom and the Netherlands are held. The A165 output is unchanged: the refetch restated no trust value or note. pnpm bench evs fetch runs the adapter, with trust kept as an alias. The memo is docs/research/agency/EVS-WVS-ITEMS.md.

The statistic is the published mean because it is what the table prints. The table gives the full distribution, the valid-answer base, the mean and the standard deviation, and no top-box share. A share answering 7 to 10 would have to be summed from category percentages whose denominator includes don't know and no answer, which is a number the publisher does not publish.

Why. Construct first, written in the memo before values were read in. Agency asks how able people are to turn an intention into action. A173 does not observe action; it observes the felt capacity to act, whether people believe what they decide changes what happens to them. That is the half of the question no current row reaches: Agency rested on new business density and two Doing Business rows frozen at 2019 (A14). It is a perception and is labelled as one, as interpersonal_trust is.

Reported as findings, not tests (D118), from a local run on dataset 7.1.0: the row covers 37 of 53 countries, fieldwork 2017 (eight countries, the United States among them) to 2023 (India), every value stamped with the release year 2022. Brazil is 7.5 (16th of 37, fieldwork 2018). The value correlates with log GDP per capita at r = -0.154 (n 36); its wealthAttribution delta is -0.059. Agency moves from r = 0.638 to 0.579 against log GDP, its mean confidence from 0.383 to 0.483, which crosses O1, and it stays scored in 52 countries. The United States, Nicaragua and Venezuela, which were Agency on the two frozen rows alone, now carry a third. The guardrail, mean confidence against log GDP, falls from 0.333 to 0.286 (n 51), together with D128. No redundant pair forms; the row correlates with interpersonal trust at -0.14.

Cost. Response style on a 10-point scale differs across cultures, so part of the ordering is how samples use a scale: Mexico, Uruguay and Colombia sit at 8.1 to 8.2, Japan at 6.0 on a mail survey. A closed or electoral autocracy can read high, and two do: Vietnam is third at 8.1 and Nicaragua fifth at 8.0, surveyed in 2019-20 after the 2018 crackdown. That is the A13 pattern in a perception, recorded as artefact A15 and not corrected. The mean is printed to one decimal over a spread of 6.0 to 8.2, so ties are common. Fieldwork years differ by up to six years and the stamped year hides it; the note carries each country's survey year. Sixteen countries have no row, Costa Rica, Ireland, Israel and South Africa among them.

Overturned by. Evidence that the cross-country ordering is mostly response style, for example anchoring-vignette or scale-use studies that reorder the frame when applied, which would make the row a measure of how samples answer 10-point questions; a behavioural Agency row with frame coverage that contradicts it, which would make it a check under D60; or a release whose pooled microdata let the three held countries in, which needs a new entry for the pooling rule.

D128: Civic participation is scored in Shared purpose from charitable membership (A080_01)

Recorded 2026-10-01. Extends D64's adapter under D117 and D118, beside D127.

Decision. civic_participation (Shared purpose) moves from gap to adapter. The value is the publisher-weighted share of respondents who mention membership of a humanitarian or charitable organisation, A080_01, in the same pinned results PDF, read by the same item table and hold rule as D127. Class C, higher_better, unit % mentioning. The definition narrows from "active membership in associations, unions, parties and community organisations" to what the row measures. Religious membership (A065) is left out because it reads religiosity; an "any non-religious membership" count would match the old definition better, but the aggregate table does not publish it and the adapter does not use microdata.

The joint file harmonises two questions. EVS 2017 shows a list and asks which organisations the respondent belongs to; WVS 7 reads each type aloud and asks whether the respondent is an active member, an inactive member or not a member. The codebook recodes both WVS member answers to 1, so the stored "Mentioned" column is belongs (EVS) or active or inactive member (WVS), over all respondents including don't know and no answer.

Why. Construct first, written in the memo before values were read in. Shared purpose asks whether people can imagine themselves as participants in a common project. Joining an organisation whose purpose is to act for people outside one's own network is a reported behaviour, and of the ten membership items it is the closest to acting for strangers; unions and professional bodies read interest groups, sports and culture read leisure, and parties are the democratic channel A5 retired. Shared purpose had two scored rows and was the thinnest dimension after Trust.

Reported as findings (D118), same run: 37 of 53 countries, fieldwork 2017 to 2023. Brazil is 9.7% (25th of 37). The value correlates with log GDP per capita at r = -0.331 (n 36); the wealthAttribution delta is -0.172. Shared purpose moves from r = 0.457 (n 47) to 0.202 (n 50) against log GDP, its mean confidence from 0.260 to 0.343, still under O1, and it is scored in 51 countries instead of 47: Nigeria, Vietnam, Singapore and Venezuela clear the coverage floor. Brazil's Shared purpose moves from 34.9 to 30.7. No redundant pair forms. The A13 fear did not materialise: China reads 2.7% and Vietnam 9.4%, so state mass organisations, which are unions and youth and women's federations, do not inflate a charitable item. Cuba is not in the release.

Cost. Question format. The eight EVS countries in the frame average 9.9% and the 29 WVS countries 19.4%; among the ten countries surveyed by both programmes the WVS share is higher in eight, by a median 1.9 points (Great Britain +16.0, Czechia -5.0), with fieldwork years also differing inside the pairs. Inside the EVS group the share tracks income at 0.83; across the frame it runs against it. Part of the spread is the format, and a reader should check that before reading the negative income correlation as evidence for the claim. Estonia's 1.5% is the floor and the value most exposed to it; Kenya (38.4%) and Indonesia (37.9%) top the row with inactive members counted. Membership is not activity. One item is narrower than the construct. The row is artefact A15 with D127.

Overturned by. Microdata showing that the EVS-WVS difference holds within countries after fieldwork year is controlled, at a size that reorders the frame, which would hold the row until a format adjustment is documented; a harmonised "any non-religious membership" or volunteering series with frame coverage, which would replace the single item; or evidence that charitable membership in some countries is a condition of employment or state programmes, which would make it the A13 case after all.

D129: Voter turnout is published as a behavioural check beside Shared Purpose, not scored

Recorded 2026-10-01. Extends D60, D118 and D121. Answers Q8 of the O1 triage sweep.

Decision. V-Dem's election turnout is published beside Shared Purpose as a behavioural check under D60 and enters no score. The check is voter_turnout in checks.ts, read from v2eltrnout in the Full+Others archive the V-Dem adapter already pins (v15, 2025-03-04, CC BY-SA 4.0). The codebook (3.1.4.3) defines it as the percentage of all registered voters who cast a vote in the national election according to official results; it is not the voting-age-population reading, which is v2elvaptrn. Where executive and legislative elections fall on the same day V-Dem codes the executive turnout, and the country-year takes the maximum across that year's elections. Unit % of registered voters, direction higher_better, tier expert_panel. It covers 52 of 53 countries; China holds no national election V-Dem codes.

V-Dem codes election variables in election years only, so the adapter's single-year design would leave most of the frame blank at 2024. Each entry in the adapter's variable table now names its year rule. release_year keeps the old behaviour for civil-society strength and polarization. latest_election reads each benchmark country's newest row up to 2024 that carries a value, emits it with that row's year, and never reaches back to an older election when the newest one is out of scale. Election years run 2016 (Haiti) to 2024: 18 countries at 2024, 15 at 2023, 9 at 2022, 7 at 2021, one each at 2020, 2019 (United Arab Emirates) and 2016. The observation note carries v2elcomvot (compulsory voting, codebook 3.1.2.3) and v2x_regime (Regimes of the World) for the same country-year, with their codebook labels, as context and never as values. CheckDef.pinned gains years, only by default or latest_up_to, and /sources prints "in each country's latest year with a value, up to 2024" for this check instead of "for 2024".

No registry gap is added. Polarization shares its id with a declared gap because that gap names a measurement Shared Purpose still wants; turnout is not such a measurement, so the check stands alone. No scoring rule is added: the triage sweep's proposed rule (score where v2x_regime is 2 or 3 and v2elcomvot is 0 or 1) is recorded in the memo and not wired.

Why. D118 chooses rows for what they measure. Turnout is a revealed act of taking part in a common decision, the most direct behaviour the benchmark can see for Shared Purpose, and that is why it is published. It is not scored because three things move it that are not whether people see themselves in a common project, and the number cannot tell them apart. It reads the democratic channel, which A5 retired voice and accountability for: a country without competitive elections reads low or high for reasons of regime. Compulsory voting turns it into a reading of the law: Brazil enforces it (code 2, sanctions enforced at minimal cost; 79.42% in 2022), and the eight countries that enforce sanctions average 82.6% against 65.4% for the 36 where voting is voluntary. And autocracies manage it, the A13 trap from the other side: Vietnam reads 95.6, Rwanda 98.2, Ethiopia 93.6 and Singapore 93.6, which is mobilisation, not participation. By regime at the election year the means are flat, 68.8 closed autocracies (n 3), 70.9 electoral autocracies (16), 67.5 electoral democracies (15) and 69.6 liberal democracies (18): the series does not separate the regimes it should, which is the construct failure in one line. The sweep's rule would leave about 27 scorable countries and exclude Brazil, the benchmark's subject, so a scored row is declined.

Cost. Shared Purpose stays at two rows, 47 published countries, mean confidence 0.260. No score moves: all 477 country-dimension cells, their confidences and every composite are identical with the check present; the only change in index.json is the new check row. Reported as a finding, not a reason: behaviouralChecks puts the published value at r = 0.053 against log GDP per capita (n 52) and 0.007 against the Shared Purpose score (n 46). Turnout is not a wealth proxy; it fails on construct alone, as polarization did. The latest-election rule mixes vintages up to eight years apart (Haiti 2016), and a country-year maximum mixes presidential and parliamentary elections across countries; the year shown on every row is the only guard. The United Arab Emirates' 34.8% (2019) is turnout of a hand-picked electoral college, not of citizens. The adapter now fully parses every benchmark row up to 2024 instead of only the release year, which costs a few seconds. The dataset version does not move: no scored row, field or country changes.

Overturned by. A scoring rule conditioned on regime and compulsion that the project adopts on construct, which would move the conditioned subset to indicators.ts as a row and leave the rest as this check; a turnout source that measures participation outside the electoral channel; a V-Dem release that changes v2eltrnout's denominator or drops coverage below half the frame; or evidence that readers take the check for a score, which under D60 would retire it.

D130: Brazil's layer publishes a computed map of Adaptability among its income peers

Decision. The Brazil layer gains /brasil/adaptacao, a Portuguese page that reads one capability for one country as a map: the score and its confidence as two numbers, each capability row the score rests on (value as published, year, position on the 0 to 100 frame, source), each condition beside it with its value, rank and the two correlations D122 already publishes (with log GDP per capita and with the dimension's score), and the country's place among its income peers. buildCapabilityMap in packages/core/src/pipeline/capability-map.ts computes all of it from the published files and nothing else; the page renders it through the lexicon (Lexicon.capabilityMap, English and Portuguese). Only Adaptability is published. The function takes the dimension as a parameter, the layer section id is map.<dimension>, and MAP_DIMENSIONS lists what is published, so a second capability is a registry line, a slug and a decision entry. The slug is adaptacao because the lexicon names the dimension Adaptação.

The peer rule: the 10 countries nearest the subject in log10 GDP per capita (NY.GDP.PCAP.PP.KD, latest published year per country), ties broken by iso3, the subject never its own peer, a country with no income never a peer. The set is the same for every dimension. A peer median is taken over the peers that have the value; a peer with no score is listed and left out of the score median. Above, below or level is decided at the precision the value is printed at (one decimal for a 0 to 100 position, three for a condition's published value). For the set to be computed from published data, income has to be published: diagnostics.json gains income, the latest GDP per capita per country with its year, rounded to the dollar. That is a field added, so the dataset moves to 7.2.0 and no published number changes.

The no-prescription rule: every sentence the page computes compares one value with one median, or states a correlation. No string in capabilityMap may say what a country should do, rank it among its peers, or explain why a gap exists. The reading lists rows above, below and level with the peer median and conditions the country has more or less of, by id, through readCapabilityMap; the words are the lexicon's. Splitting the agenda stays splitAgenda's job (D39); this page does not sort anything into raise or hold, and the FlagField it draws is the one chart (D67), fed only the peers and the subject.

On dataset 7.2.0 Brazil's peers are Colombia, Mexico, Thailand, Paraguay, the Dominican Republic, Peru, China, Vietnam, Indonesia and Argentina, at US$ 15,091 to 27,847 against Brazil's 20,025 (2025). Brazil's Adaptability is 65.4 (confidence 0.68, good) against a peer median of 70.3. All five scored rows sit below the peer median; export diversification only just (83.0 against 83.4). Fixed broadband, the one condition, is above it (24.1 against 16.1 per 100 people) and correlates 0.85 with income and 0.67 with the score.

Why. The owner's framing is that the page is the map toward adaptability, not a verdict on it. D122 made the map possible by separating what a country does from what it has; this page lays the two side by side for the reader the layer serves. WHY.md rules out a ranking, a target and policy advice, and says a national score is a coarse proxy, so the comparison is a median and the limits are on the page: the proxy, the peer rule's blindness to everything but income, correlation read across the whole set, and the row caveats of D119 and D120 (merchandise concentration is a product mix; long-term unemployment passes a gate, and Brazil's comes from PNAD Contínua, a household survey). A hand-picked peer list would be the comparison the author wanted; a fixed rule on the series the wealth tests already read is one a reader can recompute. Ten is enough for a median to mean something and few enough that the set stays near in income (a factor of about 1.4 either side for Brazil); a ±band was rejected because its count swings with where a country sits in the income distribution.

Cost. Income is the only thing peers share: size, region, production structure and regime are outside the rule, and a GDP revision can swap a peer. The page is Brazil-only and Portuguese-only, so the comparison is not yet checkable on the ground layer for another country. Level is decided at printed precision, so 83.0 against 83.4 reads as below. The row caveats are static lexicon text and must be kept in step with D119 and D120 by hand. Publishing GDP per capita puts income one click from every score, which the project has so far kept inside correlations.

Overturned by. A reader or reviewer showing that a sentence on the page is read as advice despite the rule, which would cut the reading section back to numbers; evidence that the peer set is unstable across releases (more than half the peers swapping on a routine re-ingest), which would move the rule to a wider set or a band; or a second dimension's map showing the layout does not carry it, which would move the page to the ground layer with the layer as one reading of it.

D125: The GEM extension is held, because it adds evidence mostly where income already is

Recorded 2026-10-01. Brief: the Experimentation section of docs/research/O1-TRIAGE-SWEEP.md. The extraction is kept in docs/research/experimentation/GEM-AND-DESIGNS.md and gem-extract.py.

Decision. early_stage_entrepreneurial_activity and failure_tolerance stay on their original 16 countries. The values for every other benchmark country in the seven GEM Global Reports for survey years 2019 to 2025 were read, checked against the printed pages and recorded in the memo, and they are not entered in manual.json. Industrial designs (D126) ships without them.

Why. The guardrail beside O1 says confidence must not come to track wealth (D117). GEM participation is chosen and paid for by national teams, and the 14 benchmark countries it has not surveyed since 2019 are, apart from Singapore, lower-income: Vietnam, the Philippines, Malaysia, Nigeria, Kenya, Rwanda, Ethiopia, Bolivia, Paraguay, Honduras, Nicaragua, Cuba and Haiti. Measured on dataset 7.2.0, where the guardrail is 0.286: the extension over 2022 to 2025 raises it to 0.357; widening the window to 2019 to 2025, to reach poorer countries surveyed earlier, raises it further to 0.387, because three of the five countries the wider window adds are high-income. With industrial designs alone it is 0.298. Experimentation's own confidence against log GDP would go from 0.20 to 0.57. The extension would raise Experimentation's mean confidence from 0.225 to 0.350, and that gain would sit almost entirely in countries the benchmark already knows best.

The construct is not the problem. TEA counts people trying and fear of failure is the attitude the dimension names; the rule for a future entry stands: survey year, not report year; the latest year in the window, never an average; and only the fear of failure question GEM introduced in 2019, over adults who see good opportunities.

Cost. Experimentation stays the thinnest dimension, at mean confidence 0.27 with designs, and 37 countries stay scored on patents, trademarks and designs alone (A1). Values that were read and verified stay unpublished.

Overturned by. GEM surveying enough of the missing lower-income countries that the extension no longer raises the guardrail, or an open GEM data release that covers them; or a decision that a source raising confidence in richer countries is acceptable when its construct is sound, which would supersede this entry and D117's guardrail together.

D126: Resident industrial design applications are scored in Experimentation

Recorded 2026-10-01. Same brief as D125.

Decision. A new Experimentation row, resident_industrial_designs_per_million, World Bank IP.IDS.RSCT: industrial design applications by residents at their national office, per million people, higher_better, class O. The per-head transform is the one resident_trademarks_per_million uses (per_million_population over SP.POP.TOTL), so the two rows are built alike. Coverage 50 of 53 (no NLD, VEN, HTI); 45 at 2021, the others between 2007 (ETH) and 2020.

Why. Construct first. A filed design is a registered attempt at a new product form. It costs less and is filed more often than a patent, which is closer to the dimension's many-small-experiments reading than patents are, and it is an output rather than a stock. Reported as findings, not tests: the probe gives r = 0.02 for the raw count against GDP; the scored row's normalised value correlates with log GDP per capita at r = 0.494 (trademarks 0.578, patents 0.524), its wealth-attribution delta is 0.007, and its r with resident_trademarks_per_million is 0.816 (n 49), under the 0.85 redundancy flag. On dataset 7.3.0 this row takes Experimentation from r = 0.623 to 0.572 with log GDP and its mean confidence from 0.225 to 0.271; the guardrail moves from 0.286 to 0.298. The GEM extension it was built beside is held (D125).

Cost. Three traps, all in the registry note. China subsidised design filings as it did patents, so its count runs ahead of the attempts behind it; it reaches the fence with Korea, Turkey, Germany, France, the United Kingdom and Switzerland. EU applicants increasingly file at the EUIPO, which a national resident count misses, so EU members read low. The Netherlands files through the Benelux office and has no national series. The row partly duplicates trademarks (r 0.816), so filing culture weighs twice.

Overturned by. A redundancy reading at or above 0.85 with trademarks, which would make one of the two rows a check; evidence that subsidised or strategic filing, not attempts, sets the cross-country order; or a WIPO series that counts a country's residents at every office it files at (national, EUIPO, Hague), which would replace this one.

D131: Government compliance with the courts is scored in Trust; the rest of the V-Dem sweep is declined

Recorded 2026-10-01. Answers Q5 of the research roadmap. Extends D57, D118 and D121. One combined entry: one row wired, every other candidate declined. The triage table is docs/research/vdem-sweep/TRIAGE.md.

Decision. A new Trust row in the institutional family, court_compliance: V-Dem's compliance with judiciary (v2jucomp, codebook 3.8.1.11), "how often would you say the government complies with important decisions by other courts with which it disagrees?", read as v2jucomp_osp, the measurement-model estimate on the original 0 (never) to 4 (always) scale, from the Full+Others v15 archive the adapter already pins, year 2024. 53 of 53 countries. Direction higher is better, class O, tier expert_panel. The adapter reads it as a fourth entry in its variable table under the release_year rule. It fills no existing gap: it is neither public confidence (institutional_trust) nor court throughput (court_case_clearance).

Twenty other variables were triaged for Trust and Coordination and none is wired, as a score or as a check. Coordination gains nothing.

Why this one. Trust's high end is strangers cooperating on the strength of the rules. A ruling a stranger obtains is worth something only if the state obeys the rulings it loses, so whether the rules bind the strongest party is the institutional precondition of the dimension. The variable is an expert code, the same kind of instrument as the WGI composites D23 retired, and it is even one of their inputs (through V-Dem's judicial constraints and liberal component indices, which the WGI rule of law estimate uses). What separates it is the object coded: one act whose instances are public, a court ruling against the government and the government's response, so a coder answers about a frequency that the record can check. The retired composites, and the V-Dem items this sweep declined in their place (v2clrspct impartial administration, v2cltrnslw predictable enforcement, v2exbribe, v2excrptps and v2jucorrdc bribery), ask for a characterisation of how clean or impartial a country's institutions are, or about hidden acts no coder observes: a reputation, whatever the method.

The A13 test was written before the values were read: where courts never rule against the state there is nothing to disobey, and a closed regime could read well on silence. It does not happen. Closed autocracies average 0.72 on the 0 to 4 scale and electoral autocracies 1.67, against 2.75 for electoral and 3.53 for liberal democracies; China reads 0.21, the United Arab Emirates 0.59, Vietnam 1.25. Singapore (3.54, 12th of 53) is the one autocracy that reads high, and the registry note names it. Scored, the row moves Trust against the trap the dimension already carried through bribery reticence (D123): China falls from 88.7 to 67.5, Rwanda from 78.9 to 63.9, El Salvador from 69.8 to 51.1.

Why not the others. Coordination's best candidate, range of consultation (v2dlconslt), observes who is heard while policy is made, not whether independent actors then act together; it is a deliberative-democracy component and it fails A13 (closed autocracies 0.47 above electoral autocracies -0.32; Vietnam level with Uruguay). Merit appointment, bureaucratic pay and fiscal source are properties of the state apparatus, conditions under D122, and the model starts no expert-coded condition here. The CSO consultation and participation items feed v2x_cspart (r 0.90 and 0.88). Common-good justification puts Cuba second of 53. Compliance with the high court duplicates the row chosen (r 0.95). Each verdict is in the triage table.

Findings, reported and not used to decide. The row's r with log GDP per capita is 0.533; its largest r with another Trust row is 0.36 (bribery incidence). On dataset 7.3.0 it moves Trust's r with log GDP from 0.571 (n 49) to 0.671 (n 51), a wealth-attribution delta of 0.155, close to bribery incidence's 0.162; Trust's mean confidence from 0.311 to 0.347, still under O1's 0.40; its scored countries from 50 to 52 (the United Arab Emirates and Haiti publish for the first time, both on the institutional family alone; Cuba stays below the floor); and Brazil's Trust from 41.4 to 55.9 at confidence 0.396. The guardrail does not move: the mean confidence across dimensions correlates with log GDP per capita at 0.298 before and after, because the row covers every country.

Cost. The row reads regime: r 0.88 with V-Dem's electoral democracy index. That follows from the construct, since in this frame the states that do not obey their courts are autocracies, but it means Trust now carries a democracy signal, and r 0.82 with the Coordination row civil_society_strength from the same source. Trust's income correlation rises by 0.10 toward the 0.70 line O2 reports. Two countries publish Trust on one family, which D57 counts in familyBalance (15 countries now, 13 before). An expert code is still an expert code: the source tier is expert_panel and the note says to read it as a judgement of a public act. The dataset version is not bumped in this change; adding a scored row is a minor bump when it ships.

Overturned by. Evidence that coders score compliance from a country's general reputation rather than from rulings and responses, for example a validation against documented non-compliance events that the code does not track, which would make it the reputation D23 retired and return it to a check or retire it; a closed regime reading high on silence in a later release, the A13 failure written down here; a harmonised court-performance series (clearance or enforcement) across the frame, which answers the institutional family more directly and would make this row a check beside it; or a redundancy reading at or above 0.85 with another Trust row.

D132: Confidence in the courts is published as a check beside Trust; institutional_trust stays a gap

Recorded 2026-10-01. Extends D57, D60, D64, D118 and D121. The memo is docs/research/trust/EVS-WVS-INSTITUTIONAL-TRUST.md.

Decision. The Joint EVS/WVS adapter reads a fourth item from the same pinned results release (v5.0.0, weighted by gwght): E069_17, confidence in the justice system and courts. It is emitted as the behavioural check __check__institutional_trust, declared in checks.ts under the gap's own id with ingest: 'adapter' and a pinned entry naming ZA7505_cdb_Tables.pdf, as polarization shares its id with its gap (D121). The stored value is the published share answering "a great deal", over all respondents; the other published shares (quite a lot, not very much, none at all, don't know, no answer) and the fieldwork year are quoted in each observation's note. The same hold rule applies: Germany, Great Britain and the Netherlands have separate EVS and WVS rows and are held. 37 of 53 countries, fieldwork 2017 to 2023. The institutional_trust row stays ingest: 'gap'. Nothing is scored.

Why this item. The construct was written before any value was read. Courts answer the dimension most directly: a court is where a stranger goes when the rule is broken, so confidence in it is the respondent's estimate that the rule will be enforced. Civil service was the second choice. Government, parliament and parties were argued out because they read the incumbent and the electoral cycle of each fieldwork year, not whether the rules hold whoever governs; police was argued out because it carries crime exposure, which D44 showed travels with income. The table prints no mean and no "great deal plus quite a lot" column, so the one published number that reads confidence by itself is "a great deal"; the sum would be computed from two rounded columns, which D64's rule forbids.

Why a check and not a score. The A13 test was run before deciding, on V-Dem v15's 2024 Regimes of the World (v2x_regime) read from the pinned archive. Share saying a great deal: closed autocracies 28.1 (China, Vietnam), electoral autocracies 23.5 (n 9), liberal democracies 14.0 (n 11), electoral democracies 7.9 (n 15). Counting quite a lot too: 88.3, 60.3, 60.9 and 36.2. India (39.7), the Philippines (33.0) and Indonesia (30.6) lead the published share; Vietnam and China lead the combined one. Every other item tested reads the same way, autocracies above democracies on both statistics: civil service (great deal 16.9 against 5.7), police (23.6 against 15.9), parliament (18.4 against 4.2) and government (26.2 against 6.5). The item also does not track the institutional fact it would stand for: across the 37 countries it correlates -0.13 with V-Dem's government compliance with the courts (court_compliance, D131), the row that does score in this family. Scored, it would rank highest the states whose courts are least able to rule against them. That is A13's failure, and D60 says the honest form is a check.

Findings, reported and not used to decide. The published share correlates -0.18 with log GDP per capita (n 36); the combined share would be 0.30. behaviouralChecks reports r -0.056 with the Trust score (n 37). Brazil reads 11.8 percent a great deal (38.5 quite a lot), 19th of 37. Trust does not move: r with log GDP 0.671 (n 51), mean confidence 0.347, 52 countries scored, Brazil 55.9 at confidence 0.396, all as on dataset 7.4.0. The guardrail, the mean confidence across dimensions against log GDP per capita, stays 0.298. The A165, A173 and A080_01 observations are unchanged in value and note; only their retrieval stamp moves.

Cost. Trust still misses O1's 0.40 confidence target and the institutional family still has no survey row: this change publishes evidence and closes nothing. A reader may take the check's high values for autocracies as a finding; the attached note is the only mitigation. Storing the "a great deal" share alone reads only the top category, which is a narrower and noisier statistic than the conventional two-category share; the alternative needed a computed value. No dataset version bump: a check adds no scored row.

Overturned by. A release that publishes a valid-answer mean or a two-category share for the battery, with the regime pattern gone (autocracies no higher than democracies) when tested the same way; pooled microdata with a list-experiment or anonymity adjustment that removes the deference component in closed and electoral autocracies; or a confidence measure that tracks court_compliance (r at or above 0.5) rather than running against it, which would show it reads the institution and not the regime.

D133: Brazil's layer publishes the capability map for all nine dimensions

Recorded 2026-10-01. Extends D130, which published Adaptability alone and named "a second dimension's map" as a test of the layout.

Decision. MAP_DIMENSIONS is every dimension. Each is computed by the same buildCapabilityMap, with no branch per dimension: a dimension with no conditions (Coordination, Trust, Experimentation and Shared purpose on dataset 7.4.0) returns an empty conditions list, and its page draws no conditions panel and says in one sentence that none is published. A score below the coverage floor renders through DimensionScore with the observed count, and thin confidence is read through isThinEvidence; no page holds a threshold.

Navigation. Nine tabs do not fit the layer's band and the tree stops at four levels (D73, D80), so the layer has one map section, labelled Mapa, at /brasil/mapa. That page lists the nine in the model's order with the score, the confidence, the peer median and above, below or level, each linking to /brasil/mapa/<segment>. The pages under it are not nodes: the Mapa tab owns them by address prefix, so the band lights Mapa on each. The segment is mapSlug of the pt-BR lexicon's dimension name (Adaptação is adaptacao, Propósito compartilhado proposito-compartilhado), so the lexicon is the one place a name is declared; a test pins the nine segments, so a rename that would move a published address fails until a redirect is added. /brasil/adaptacao answers 301 to /brasil/mapa/adaptacao. The section id map.<dimension> of D130 is retired in favour of map.

Artefacts. Each page names the known artefacts that bear on its dimension by id, linked to /limits. The join is one table, ARTEFACT_SCOPES in packages/core/src/model/artefacts.ts: an id, the dimensions it touches or all for a structural one (A8, A10), and an optional country scope, which keeps A2 (India) off Brazil's pages. A test reads the headings of docs/KNOWN-ARTEFACTS.md and fails when the table and the document disagree. The map carries the result as artefacts, so a page holds no id.

Hand-written facts. D130's long-term unemployment caveat said Brazil's ILOSTAT value comes from PNAD Contínua, which the published output does not carry (the pinned observation note does). It is the only one. The construct caveat is now country-neutral and the Brazil sentence moved to capabilityMap.countryRowFacts, keyed by iso3 and row, with a comment naming D120 as the decision whose supersession makes it stale. No other hand-written fact about Brazil was added: every number on the nine pages is computed.

On dataset 7.4.0, against the 10 income peers of D130 (every peer scored on every dimension), Brazil reads above the peer median on Anticipation (45.8 against 42.7, confidence 0.455), Agency (55.8 against 53.9, 0.555), Coordination (86.4 against 68.6, 0.373, thin), Trust (55.9 against 53.2, 0.396, thin) and Experimentation (25.6 against 14.4, 0.430, thin), and below it on Learning (28.0 against 35.8, 0.539), Adaptability (65.4 against 70.3, 0.678), Building (28.2 against 40.1, 0.568) and Shared purpose (30.7 against 44.3, 0.433, thin). No dimension is below the coverage floor. These are findings, recorded to date the release; the pages compute them.

Why. The owner's framing for the phase is to use the instrument: thin dimensions are published findings to read and argue with. One map was a demonstration; nine is the instrument, and holding eight back would make Adaptability look like a chosen exhibit. D130 already made the function dimension-agnostic, so this is a navigation and copy change, not a model change. One section with an index keeps the four-level tree and gives the reader a single place where all nine sit side by side, in a fixed order that is not a ranking. Slugs from the lexicon avoid a second list of names; the pin keeps that from breaking links silently. Artefact ids in a table rather than in page prose mean a new artefact, or a dimension it touches, is one line and is checked against the document.

Cost. Four of the nine pages read thin evidence, and Coordination's above-median reading rests on 0.373 confidence and three rows (A3, A9, A12): a reader can quote it as a finding. The pages say thin plainly and name the artefacts, which is the mitigation and not a fix. The index puts nine above-or-below words in one column, which invites counting them: five above and four below is not a verdict on Brazil, and the page says the capabilities do not add up. Confiança is both the Trust dimension and the confidence label in Portuguese, so the Trust page carries the same word twice in different senses. The artefact pages are English. The slug of a dimension now depends on lexicon copy.

Overturned by. A reader reading the index as a scorecard (counting above and below as a result), which would drop the position column from the index and leave it on each page; a lexicon rename the redirect table cannot absorb; or a second country layer, which would move the map index to the ground layer as D130's clause already anticipates.

D134: Mexico, Colombia, Chile and Argentina get Spanish layers with the computed sections

Recorded 2026-10-01. Supersedes the sentence "Brazil is the first and today the only layer" in D69, and answers the second-layer clause of D130 and D133 for these four countries.

Decision. The owner set the phase as using the instrument, and chose four Spanish-speaking countries to receive a layer each, like Brazil's: Mexico at /mexico, Colombia at /colombia, Chile at /chile and Argentina at /argentina. Each holds three things and nothing else: an overview, the agenda at /<slug>/agenda and the capability map at /<slug>/mapa with nine pages under it. All four read one lexicon, packages/core/src/i18n/es.ts, in neutral Latin American Spanish.

The four are entries in COUNTRY_LAYERS, not folders. One dynamic [layer] route serves every layer without a folder of its own; Brazil's static /brasil folder wins over it and is unchanged, and any other first segment answers 404. The map pages of every layer, Brazil's included, render one shared component through the layer's lexicon and LAYER_WORDS, so Brazil's nine map pages and the Spanish ones cannot diverge in structure.

What the Spanish layers do not hold, and why:

• No institutions section. data/institutions has Brazil alone, and INSTITUTION_MAPS stays ['BRA'].
• No subnational section. No subnational data exists for the four.
• No support page. /brasil/apoie names Brazilian funding venues; the ground-layer /support is the comparative page, and a Spanish one would be a page of claims the project has not done the work for (D71, D78).
• No hand-written overview. Brazil's overview is prose about Brazil. The Spanish overview states nothing the published files do not: the nine scores with the confidence beside each, in the model's order, links to the map and the agenda, and the known artefacts that bear on the country's dimensions, by id, from ARTEFACT_SCOPES. No sentence on it is a claim about the country.

Terms. The confidence number is "solidez de la evidencia", "solidez" for short, because "confianza" is the Trust dimension; the pt-BR copy made the same move to "solidez da evidência". The score is "puntuación" throughout, never "nota". Two dimension names are chosen for a Spanish reader rather than copied: Agency is "Iniciativa", because "agencia" reads as an organisation, and Building is "Ejecución", because "construcción" reads as the construction sector. The nine map segments are mapSlug of the Spanish names (anticipacion, iniciativa, coordinacion, confianza, aprendizaje, experimentacion, adaptacion, ejecucion, proposito-compartido), pinned by a test in es.test.ts. Numbers format as es-419.

Rendering. A lexicon may now name the countries it is written for (Lexicon.layerCountries). pnpm bench agenda renders {ISO3}.es.md for the four only, and bench institutions writes no Spanish feed for Brazil. The viewer checks at load that the lexicon's list and the Spanish entries in COUNTRY_LAYERS name the same countries. Portuguese keeps rendering every country, which is the D69 cost this does not reopen. The feed reads the Spanish agendas through agendaHrefInLanguage, which already gates on the layer, and the sitemap lists every new page from the registry.

Hand-written facts. Two, both in capabilityMap.countryRowFacts against D120, on the long-term unemployment row: Argentina's ILOSTAT series is the urban-only Encuesta Permanente de Hogares and the ILO flags the value unreliable, and Mexico's series fails the plausibility gate in every year, so the row is empty. The published output carries neither the survey name nor the reason a value is missing. Colombia and Chile get none: both are labour force surveys that pass the gate.

On dataset 7.5.0 against the 10 income peers each (every peer scored on every dimension): Mexico reads above the peer median on seven dimensions and below on Coordination (67.8 against 69.6, 0.373, thin) and Trust (44.5 against 60.1, 0.387, thin). Colombia reads above on Anticipation, Agency, Experimentation and Shared purpose and below on the other five. Chile reads above on six and below on Trust (66.7 against 67.4), Adaptability (64.9 against 71.0) and Building (32.3 against 37.2). Argentina reads below the peer median on all nine. Peers: Mexico THA BRA DOM CHN COL PRY ARG CRI PER CHL; Colombia BRA PRY MEX PER THA VNM IDN DOM ECU CHN; Chile URY CRI ARG MYS TUR PAN CHN DOM EST THA; Argentina CRI CHL CHN DOM URY THA MEX MYS TUR PAN. These are findings that date the release; the pages compute them.

Why. D69 grew a layer only where country-specific work was done, and argued that its existence was evidence of the work. D130 and D133 made the capability map a computed reading that needs no country-specific work: the same call, a lexicon and a peer rule. A layer that holds only computed sections is the instrument used, which is the phase the owner set. Keeping the overview computed keeps D69's argument honest: the Spanish layers claim nothing a reader cannot recompute, and where the project has done no work (institutions, states, funding) they publish no page. One route from the registry means a fifth layer is one entry and a lexicon, never a copied folder that drifts.

Cost. The Spanish layers are thinner than Brazil's and look like it: an overview of nine rows is not what /brasil is. One neutral Spanish is not the Spanish of any of the four, and es-419 prints decimals with a point, which Argentina, Chile and Colombia write with a comma. Argentina reads below its peer median on all nine dimensions, and a column of nine "below" invites the scorecard reading D133's overturn clause names. The ground-layer map index D130 and D133 anticipated for a second layer is not built: five countries have a map and 48 do not, and the map stays reachable only through a layer. The artefact pages, method, glossary and limits stay English, as for Brazil. "Iniciativa" and "Ejecución" diverge from the English names, so a reader moving between layers meets two words for one capability.

Overturned by. A Spanish reader showing that a sentence on these pages reads as advice or as a claim about the country the data does not carry; the four layers' sections staying identical to each other through the next releases while Brazil's grow, which would show D69 was right that a layer without country work is a translation, and would move the map to the ground layer for every country with the layers as readings of it; or a reader in one of the four countries finding the neutral Spanish wrong enough to need a national lexicon.

D135: The evidence corpus is complete when every grid cell is closed, and a cell can close without a record

Recorded 2026-10-01. Extends D33 and D76.

Decision. "Exhaustive" evidence research has one meaning: every cell of a grid of the 53 registry countries against the declared gaps a delivery can evidence is closed. The columns are EVIDENCE_GRID_INDICATORS in packages/core/src/model/research.ts: large project delivery, institutional responsiveness, disaster preparedness, public-private collaboration, university-industry collaboration, government foresight, regulatory sandbox activity and adult learning participation, 424 cells. A cell closes with an evidence record, or with a no-case note in data/evidence/searched.json: the date, the candidate lists searched, and each candidate with the first inclusion test it failed. A record always outranks a note. buildEvidenceGrid in packages/core/src/pipeline/research.ts derives the grid, the D76 inventory carries it and drops closed cells from its queue, and pnpm bench validate prints it on every run. The validator errors when a column stops being a declared gap and warns when a record has made a note stale.

Why. The corpus could only grow, never finish. D76's slot queue listed every uncovered country-gap pair, about 900 of them, including survey constructs a delivery cannot evidence, and a slot only left the queue when a record filled it. A country with no sandbox or no foresight unit would stay open forever, and the queue would keep sending research to it, which is the pressure that turns a weak case into a record. A dated note makes the negative result a finding that can be checked and repeated. The first run of the check found that civic_participation had already been moved to a measured row from the Joint EVS/WVS release, which is why it is not a column.

Cost. A no-case note is only as good as the lists searched. A researcher who searches a thin list closes a cell too early, and the grid then reads as complete when it is not. The column set is judgment: adult learning participation and university-industry collaboration sit close to survey and dataset constructs, and records against them are often one step from the indicator's definition. The grid counts closure, not quality, so a cell closed by a weak record looks the same as one closed by a strong one.

Overturned by. Notes that a later search with the same lists overturns at a high rate, which would show the notes are closing cells on thin searches and need a second researcher; or a comparable series for any column, which promotes the gap under D20 and removes the column.

D136: Every country gets the capability map in the ground layer, and the layers' maps become readings of it

Recorded 2026-10-01. Answers the overturn clauses of D130 ("would move the page to the ground layer with the layer as one reading of it") and D133 ("a second country layer, which would move the map index to the ground layer"), and the cost D134 recorded: five countries had a map and 48 did not. Replaces the per-lexicon countryRowFacts of D133 and D134.

Decision. The ground layer publishes the capability map for all 53 countries, in English: /country/<ISO3>/map, the index of nine capabilities with the score and the confidence as two numbers and the position against the income-peer median, and /country/<ISO3>/map/<dimension>, one page per capability. The segment is the dimension id, the one /capabilities/<id> uses, so the English addresses depend on the registry and not on lexicon copy. Both pages render the same two components the layers render, CapabilityMapIndex and CapabilityMapDimension, through EN. What used to be a layer argument is now a MapReading (apps/web/src/lib/map-reading.ts): the iso3, the language, the addresses the pages link between and the decisions they cite. groundMapReading and layerMapReading build the two kinds, so a layer's map and the English one cannot diverge in structure and the layers' pages render as before.

Navigation. The map is one page of the English reading, labelled Map, after Agenda, in the order a layer holds its sections. The nine capability pages are owned by the prefix, as /brasil/mapa owns its nine, so the tree stays at four levels: Countries, the country, the reading where a layer exists, then Map. On a map page of a country with a layer, the crumb that offers both readings points the other reading at the same page there (mapCounterparts), so switching language keeps the reader on the capability; elsewhere the crumb is unchanged. Peer flags on the field chart link to the peer's profile, as they do in the layers.

Hand-written facts. The layers carried three sentences about one country's long-term unemployment row (Brazil's PNAD Contínua, Argentina's urban EPH and ILO flag, Mexico's empty row), each written per language. The ground layer shows such facts, so they move to one keyed table, COUNTRY_ROW_FACTS in packages/core/src/model/row-facts.ts, read by every reading: a country, a row, a kind, the survey's own name where the kind needs it, and the decisions it rests on. A lexicon holds one template per kind (capabilityMap.rowFacts), never a sentence per country; the gate's floor in the Mexican sentence is read from LTU_GATE. buildCapabilityMap carries the country's facts as facts. On a ground layer that reads every country, a fact shown for a country with a layer and withheld from another with the same fact would be a selection, so the table holds every country the pinned ILOSTAT release says it about: household surveys for Brazil, Honduras and Paraguay; a household survey the ILO flags unreliable for Kenya; an urban-only survey the ILO flags for Argentina; an ILO unreliable flag for the United Arab Emirates, Finland and France; and no year passing the D120 gate for South Korea, Mexico, Peru, the Philippines, El Salvador and Uruguay. 14 entries. A test reads data/observations/ilostat-ltu.json and fails when a survey or a flag in the file and the table disagree, and another pins the Portuguese and Spanish sentences as they were, so the layers' text is unchanged.

Deriving these from the published data was considered and declined for now. data/out does not carry the observation note: the survey and the flag would need a new published field (a dataset minor, a rescore and a schema change), and the gate's held countries are in no observation at all, so the empty-row fact would also need the adapter to publish its drops, which means a re-fetch.

Degradation, on dataset 7.5.0. Cuba and Venezuela have no GDP per capita in diagnostics.income, so their maps form no peer set: every position reads "no comparison", and the pages say in one sentence why (capabilityMap.noIncome, new in all three lexicons). Cuba is below the coverage floor on Anticipation, Agency, Coordination, Trust and Shared purpose, Haiti on Shared purpose; those pages render through DimensionScore with the observed count. Rwanda's and Ethiopia's medians on two dimensions rest on nine scored peers of 10. Every other map has 10 scored peers. Argentina is the one country below the peer median on all nine; no country is above on all nine.

The English lexicon now gives six names an article inside a sentence ("the United States", "the Netherlands", "the United Kingdom", "the United Arab Emirates", "the Philippines", "the Dominican Republic"), which the map's headings need and the English agenda already lacked; the six English agenda documents are re-rendered. No published field is added, so the dataset stays 7.5.0. App 1.22.0.

Why. The owner's phase is to use the instrument, and the instrument was reachable only through five layers. D130 and D133 made the map a computed reading that needs no country work, and D134's overturn clause names exactly this move: the map belongs to the ground layer for every country, with the layers as readings of it. WHY.md asks whether capability is separable from wealth; a reading per country against the countries at the same income is the form of that question one country's reader can check, and it should be checkable for any of the 53, not for the five with a language of their own. The ground layer is English because it is the benchmark itself, not a translation of it (D69). Dimension ids rather than English slugs keep one address convention across /capabilities and the map. One fact table with templates is the smallest change that stops three languages carrying three copies of country prose that drift apart.

Cost. 530 new pages, each a reading a reader can quote: nine above or below words per country, and for some countries (Argentina among them) a column that invites the scorecard reading D133 names. The index keeps the mitigations the layers have (no verdict colour, no counts, the note that the capabilities do not add up); it is not a fix. The fact table is still hand-kept: a re-fetch that moves a held country out of the gate is not caught by the test, only a moved survey or flag is. The map for Cuba and Venezuela is a table of scores without the comparison that defines the page. The crumb switch lands on the matching page only for the map; the agenda still switches to the other reading's front page. Five tabs on Brazil's English reading wrap to a second line at phone width.

Overturned by. A reader treating the English index as a scorecard (a ranking of a country's capabilities or a count of above and below quoted as a result), which would drop the position column from every index, layers included; a published field carrying the survey and the gate's drops, which would retire COUNTRY_ROW_FACTS in favour of the data; or the fact table growing past one indicator, which would show the facts are a data problem and move them into the adapter's output.

D137: The one-factor test is computed and published every release

Recorded 2026-10-01. Puts in the data the test docs/WHY.md names first under what would show the idea was wrong: "the nine dimensions collapse into one factor across a wide country set, with no information beyond income per head." Until now it was computed by hand, once, on a different basis.

Decision. diagnostics.factorStructure publishes, on every bench diagnose, a principal-component reading of the nine dimension scores:

• The correlation matrix of the published dimension scores (score, never blendedScore, so the panel cannot enter) over the complete cases: countries with all nine scored. Every dropped country is named with the dimensions it lacks. Pairwise correlations over different country sets were rejected because the result need not be a correlation matrix of anything and its eigenvalues can go negative. Nothing is imputed (the rule since D45).
• If complete cases fall below FACTOR_MIN_COUNTRIES (30), the same solution is also computed on the dimensions scored for at least FACTOR_NEAR_FULL_COVERAGE (90%) of countries, and the page names them. A one-line reading quotes that fallback only then.
• Eigenvalues by cyclic Jacobi rotation in pipeline/stats.ts, no new dependency. The first-factor share is the largest eigenvalue over the number of dimensions. Loadings are the eigenvector times the root of its eigenvalue (each dimension's correlation with the factor), signed so they sum positive.
• The first factor's country scores (standardised scores times the unit eigenvector) correlated with log GDP per capita, with its n, and r squared as the share of the factor income accounts for.
• A chance level: the first-factor share of n by p independent standard normal draws, 2000 samples, seed 20261001, mean and 95th percentile. At n 51 and nine dimensions chance alone gives about 0.19, so a share is never read against 1/9.
• Reading: at an absolute r of 0.8 or more with income, every surface says the shared factor looks like income; from 0.5 it says income accounts for part of it; below, that it is mostly something else (FACTOR_INCOME_BANDS). The sentences are templates over the numbers in apps/web/src/lib/factor.ts and report.ts, never fixed prose.

The release history is data/out/factor-history.json, written by bench diagnose from git: for every Dataset X.Y.Z changelog release, the last commit whose data/out/index.json carried that version, run through the same function. Income comes from that commit's diagnostics (D130 on) or, before that, from the GDP context series in its worldbank.json. Past output is read as published rather than rescored, because the question is what each release said. The current release comes from the run itself. Without git the committed file is kept. It sits in data/out, not data/research, because the viewer reads it and data/research holds the evidence corpus.

On dataset 7.6.0: 51 complete cases (Cuba, Haiti dropped), share 0.529 against chance 0.189 (95th 0.215); the factor correlates 0.86 with log GDP per capita (n 50), r squared 0.74. The shared factor looks like income, and the page says so. History: 0.604 at 6.2.0 and 0.498 at 7.0.0, both on 33 complete cases, when the stock rows left the scores (D122). The hand figure quoted before (0.62 to 0.50) was on eight dimensions with Trust left out and 46 countries; this rule replaces it.

Why. It is the test the project names as the one that would show it is wrong, and a test computed once by hand cannot fail in public.

Cost. A principal-component share is one summary of a 9 by 9 matrix: it says how much moves together, not what the remaining variation is worth. At 51 countries a share or an r moves by several hundredths on a handful of countries (A8). The chance level assumes Gaussian independent dimensions; bounded 0 to 100 scores are not Gaussian, and the baseline is a reference, not a significance test. Complete cases drop the least-measured countries, which are not a random sample. The 0.8 and 0.5 reading bands are judgment.

Overturned by. A country set wide enough that the factor's r with income can be estimated within a few hundredths, which would replace the bands with a confidence interval; a parallel-analysis or permutation baseline built on the actual score distributions, if it moves the chance level by more than 0.02; or an imputation rule adopted elsewhere in the model, which would make complete cases the odd one out.

D138: What is left after income is tested in aggregate every release

Recorded 2026-10-01. Follows D137, which found that the shared factor of the nine dimensions looks like income. Respects D65 and D68: every number this decision publishes is an aggregate over countries, and no residual of any one country reaches a page, a feed or a published file it was not already in. The reading rules below were written and committed before any of the four tests was run on the data.

Decision. diagnostics.residualStructure publishes, on every bench diagnose, four tests on the wealth residual of D68: each dimension's score minus the score its income predicts, from the same per-dimension ordinary least squares on log10 GDP per capita over every country scored on that dimension (unrounded, linearFit in pipeline/stats.ts). Tests (a), (b) and (d) read the complete cases: countries with all nine residuals. Nothing is imputed. Each test is built so that it can fail, and pure noise around income reads as no structure.

• (a) Is there structure left after income? The correlation matrix of the nine residual columns over the complete cases, its eigenvalues and the first factor's loadings. The first-factor share is read against the D137 chance level (firstFactorChance, 2000 draws, seed 20261001) at one fewer country, because the income fit uses up a degree of freedom. A within-dimension permutation baseline (each residual column shuffled across countries on its own, 2000 draws, seed 20261002) is published beside it as the check D137 named. Reading: a share above the chance 95th percentile reads structure (what is left after income moves together, so it is not independent noise around the income line); at or below it reads none.
• (b) Same income, different shape. Each residual column is divided by its standard deviation so every dimension counts the same. A country's shape is its nine standardised residuals minus their own mean, so a country above its income line everywhere has a level, not a shape. Its peer distance is the mean, over its MAP_PEER_COUNT (10) nearest complete-case countries in log GDP per capita (incomePeers), of the root mean square difference between the two shapes. The null shuffles each standardised residual column across countries (2000 draws, seed 20261003): every dimension keeps its spread, every country keeps its income peers, and any link between a country's nine residuals is broken. In plain words, it is what peer distances look like if the residuals were dealt out at random, with no country-specific shape. Published: the median observed peer distance with the null median and its 5th and 95th percentiles, and the share of countries whose peer distance exceeds the 95th percentile of their own null, with that share's own null mean and 95th percentile. No country is named. Reading: a median below the null 5th percentile reads alike (income peers share a shape, so income predicts shape); otherwise a share above its null 95th percentile reads differ (more countries sit far from their peers than random dealing produces); otherwise noise (peers differ by the amount chance gives).
• (c) Is it stable? From git, the same releases factor-history.json reads (D137), the residual of each dimension at each release from that release's published scores and income. For every pair of consecutive releases with the same country set, and for every dimension whose residuals moved by at least 0.05 points for some country between the two, the Spearman correlation of the residual order over the countries in both. Published per dimension: pairs compared, mean and minimum. Reading: every tested dimension's minimum at or above 0.8 reads stable; any tested dimension's minimum below 0.5 reads churning (that dimension's residual order follows indicator choice rather than countries, and the pages say so); otherwise mixed; no moved pair reads untested. Leave-one-out on the current release, per dimension: the largest change in the slope when one country is dropped, in standard errors of the slope, and the largest change in a dropped country's own residual (its residual against the line fitted without it), in residual standard deviations. Reading: any dimension above 1 residual standard deviation reads fragile, the condition D68 named as making the full-set fit wrong; otherwise robust.
• (d) How much of a profile is income? For each complete-case country, one minus the sum of its nine squared residuals over the sum of its nine squared deviations from each dimension's mean: the share of how far its profile sits from the average profile that its income line accounts for. Published: the mean (the headline), the median and the pooled share. A country's share can be negative when its income line points the wrong way. Reading: a mean of 0.5 or more reads most; 0.25 or more reads part; below reads little.

What the four mean for the claim. The strong claim, that capability is separate from wealth, is read from D137's band alone. The weaker claim, that countries at the same income have different capability shapes, so a profile carries information beyond income, is read:

• holds when (b) reads differ and (c) does not read churning;
• fails when (b) reads alike, or when (a) reads none and (b) reads noise;
• mixed otherwise.

Every sentence on /thesis and /diagnostics that states one of these verdicts is a template chosen by the computed reading in apps/web/src/lib/residual.ts, never fixed prose.

Amendment to (b), before the first run on data. The (b) rule above failed its own synthetic check, before it was run on the benchmark's scores. On 50 synthetic countries with a planted shape (each country leaning one way on four dimensions and the other way on five, by its own amount, independent of income) it read alike, because a median of root mean square distances shrinks when shapes vary along one direction, whatever the peers. It compared peers with random dealing when the question "do income peers share a shape?" compares peers with countries chosen without regard to income. Replaced, still before any run on data:

• The peer distance is the mean, over countries, of the mean squared shape distance to the 10 income peers, per dimension.
• alike when that mean falls below the 5th percentile of the same mean with every profile kept whole and the incomes dealt out at random (2000 draws, seed 20261004): income peers are more alike in shape than countries picked without regard to income, so income still predicts the shape.
• otherwise differ when the first-factor share of the shape columns (the standardised residuals minus each country's own mean) exceeds the 95th percentile of the same share with each residual column dealt out at random before the shapes are taken (2000 draws, seed 20261003): shapes line up along shared contrasts that random dealing does not produce, and since peers are not more alike than anyone, countries at the same income spread along them.
• otherwise noise.
• Published but not read: the peer distance against the random-dealing noise floor, and the share of countries whose peer distance clears the 95th percentile of their own noise floor (one in twenty by construction).

The synthetic checks in residual-structure.test.ts now hold the rule to four cases: pure noise around income reads none and noise; a planted shape reads structure and differ; a level shared by all nine reads structure and noise, so the weaker claim reads mixed; a shape that income dictates through a curve a straight line cannot remove reads alike. The rest of the decision is unchanged.

Results. Written after the first run on data, dataset 7.7.0; the rules above were not edited after it.

50 complete cases (Cuba and Haiti lack a dimension, Venezuela an income figure).

• (a) structure. The residuals' first factor carries 0.279 against a chance level of 0.191 (95th 0.217); the permutation baseline agrees (0.190, 95th 0.215), so D137's Gaussian chance level holds within 0.002 here. Loadings: Anticipation 0.84, Learning 0.70, Coordination 0.67, Trust 0.63, Shared purpose 0.61, Agency 0.27, and near zero on Experimentation, Adaptability and Building. What is left after income is not independent noise: a country above its line on the institutional dimensions tends to be above it on the others of that group.
• (b) differ. Mean peer distance 1.559 against 1.538 with incomes dealt at random (5th 1.464, 95th 1.605): income peers are no more alike in shape than anyone. The shape columns' first-factor share is 0.264 against 0.204 under random dealing (95th 0.232). Descriptive: the peer distance sits below the random-dealing floor (1.780, 5th 1.682), because part of what is left is a shared level; 6% of countries clear their own 95th against 5% by construction.
• (c) mixed between releases: 22 releases, 19 consecutive pairs on the same 53 countries, 23 dimension comparisons where a residual moved. Lowest Spearman 0.718 on Trust (n 36), then Coordination 0.740 (n 44), Shared purpose 0.759, Adaptability 0.777 and Agency 0.780; Building 0.984 and Experimentation 0.912 hold. None below 0.5. robust without one country: the largest own-residual shift is 0.249 residual SD (Agency), the largest slope shift 0.66 standard errors (Trust). D68's leave-one-out overturn condition is not met.
• (d) part. Mean 0.306, median 0.443, pooled 0.392.

The weaker claim reads holds. The strong claim reads, from D137, that what the nine share is mostly income.

Why. D137 settled the strong claim against the benchmark for the shared factor. If the project is to keep saying a capability profile is worth reading, the claim it can still make has to be tested in public, with rules that could have said no. Fixing the rules first, and committing them, is what makes a "holds" mean something.

Cost. (b) reads "differ" on a narrow margin (0.264 against 0.232) at 50 countries, and A8 applies to every figure: a handful of countries could move it across the line. The (b) rule was amended once, after a synthetic check and before any run on data; the amendment and its reason are above. Standardising each residual by its spread gives a thin dimension the same weight as a well measured one. The release test reads past releases as published, so pairs where only one or two dimensions moved carry most of it, and the 0.8 and 0.5 bands are judgment. Complete cases drop the least measured countries. The within-column nulls assume countries are exchangeable once income is out, which neighbours, regions and shared data sources violate. No test here can separate a real country-specific shape from measurement error that happens to be shared across a country's indicators, for example one national statistics office feeding several rows; only indicators from independent sources, or a measurement-noise model per row, could.

Overturned by. A release where (b) reads alike or noise, which flips the weaker claim to fails or mixed on every surface without an edit; a churning release reading on any dimension; a measurement-noise floor built from indicator uncertainty that puts the shape share inside it; or a wider country set on which the margin in (b) can be estimated within a few hundredths.

D139: A panel figure is compared with the indicators only on the dataset it was scored on

Recorded 2026-10-02. Resolves #41.

Decision. No artefact, page or report compares a Delphi run with indicator scores unless the run's datasetVersion matches the current dataset. Until then the run is a research note, which is what isDelphiRunForDataset already enforces in scoring: dataset 7.7.1 publishes no delphiScore in any cell. docs/KNOWN-ARTEFACTS.md quotes no panel figure, and A1, A2, A5, A7 and A9 lose the panel evidence they used to cite.

Why. The run data/delphi/latest.json points at is the 2026-08-26 in-session run: one panelist, 16 countries, no dataset version, scored against a frame that still held the perception rows D23 retired. Its estimates were anchored on those scores, so on Coordination its upward corrections (Uruguay 18.8 to 45, Brazil 15.5 to 35, Colombia 4.3 to 18) read as downward gaps against today's scores, and the 35.6 point mean gap #41 asked about measures the change of ruler. Rank agreement, which the anchor does not move, is 0.70 on Coordination, mid-pack of the nine.

Cost. The artefacts lose an independent second opinion until a reviewed panel exists, and A7 had no other evidence for Korea and Estonia reading low: it now rests on the indicator rows alone and was retitled.

Overturned by. A reviewed gateway panel scored against the current dataset version (#33), which makes the comparison readable again.

D140: Trust in strangers is scored in Trust as the share trusting a first-time acquaintance completely or somewhat (G007_34_B)

Recorded 2026-10-02. Extends D64's adapter under D117 and D118, beside D127 and D128. Closes #63. Answers Q6 of the O1 triage sweep.

Decision. willingness_to_cooperate_strangers (Trust, social family) moves from gap to adapter. It is read from G007_34_B, "trust: people you meet for the first time", in the same pinned Joint EVS/WVS results PDF (release 5.0.0), by the same item table, country mapping and hold rule as D127 and D128. Class P, higher_better, unit % expressing trust. The definition narrows from "trust in people met for the first time and in people of another nationality" to the first-time item alone. G007_36_B (another nationality) is not averaged in: one item per row, as D128 did for membership, and it reads attitudes to foreigners as much as trust.

The value is the sum of two published columns, trust completely plus trust somewhat, over all respondents including don't know and no answer. This is a narrow exception to the adapter's rule that a stored value is a number the publisher prints. It applies to this item only. The note on every observation quotes both addends and every other published share, so the sum can be checked against the table by hand.

Why. Construct first, in docs/research/shared-purpose/EVS-WVS-BEHAVIOURAL-ITEMS.md, written before values were read. Trust asks how far cooperation reaches beyond a person's own network. A165 asks about "most people", which a respondent can fill with their own circle; this item names the stranger. EVS 2017 and WVS 7 ask it in one wording on one four-point scale, and inside the dual-programme pairs the WVS share is a median 2.2 points higher, against a 66 point spread.

Why the sum. The scale has four points and the trust and distrust halves split at its natural midpoint, so the cut is not chosen the way a 7 to 10 cut on a ten-point scale would be. Both columns share one denominator, so the sum is exact to the rounding of the printed figures, within about 0.1. That is what separates it from the A173 refusal, where the publisher printed a mean and a top-box share would have been a chosen cut on a ten-point scale. The single printed column "trust completely" has no usable spread (0.1 to 9.5). The other printed column, "do not trust at all", would keep the rule but reads closer to A165 (r -0.79 against 0.66 for the sum) and flips the registry's unit and direction. The sum matches the unit the registry declared.

Reported as findings (D118), same run: 37 of 53 countries, fieldwork 2017 to 2023, range 7.8 (Ecuador) to 73.9 (Sweden); Brazil 22.7, 22nd of 37. The row correlates with log GDP per capita at r = 0.34 (n 36), with A165 at 0.66, under the 0.85 redundancy flag, and no redundant pair forms. Autocracies do not inflate it: closed 22.2, electoral autocracies 23.7, electoral democracies 21.5, liberal democracies 40.4. Trust's mean confidence moves from 0.347 to 0.418, which clears O1, its observed rows from 3.6 to 4.3, and its r with log GDP from 0.675 to 0.606 (n 51). Scored countries stay at 52, because no country is added. The social family now has both its rows observed. The guardrail, mean confidence across dimensions against log GDP per capita, moves from 0.298 to 0.278 (n 51). The one-factor share (D137) moves from 0.531 to 0.523, its correlation with income from 0.858 to 0.854, and Trust's loading from 0.804 to 0.760. Every D138 reading is unchanged: structure, peers differ, stability mixed, leave-one-out robust, income explains part, the weak claim holds. The largest Trust moves are falls where "most people" read high and a stranger reads low: Japan 79.4 to 64.3, Singapore 84.6 to 70.8, South Korea 80.8 to 67.6, China 67.6 to 55.8 (20th to 28th of 52) and Malaysia 68.2 to 56.7. Ethiopia rises 41.9 to 45.6 (46th to 36th) and Sweden 87.9 to 90.3. Brazil moves from 55.9 to 49.3.

Cost. The rule that a stored value is printed by the publisher now has an exception, and an exception invites a second. A summed value carries the rounding of two cells instead of one. It is a perception, class P, the same class and source as A165, so the social family is still two survey items from one release and no behaviour. The 13 countries outside the release and the three held (DEU, GBR, NLD) keep three Trust rows where the 37 now have five, so Trust's confidence gap between them widens. Ethiopia's 47.8 (fifth of 37, fieldwork 2020) is the value to check first. The EVS countries in the frame are all European and average twice the WVS ones, which is region and cannot be separated from format in the aggregate table.

Overturned by. A published single column or mean for G007_34_B in a later release, which would replace the sum and close the exception; microdata showing the EVS-WVS difference reorders the frame once fieldwork year is controlled; or a fieldwork review showing the Ethiopia sample, or any other outlier, is not national, which would hold that country's value.

D141: Vocational share, labour force participation and transmission losses move to the conditions layer

Recorded 2026-10-02. Extends D122 to three of its six Tier B rows. Issue #68. Dataset 8.0.0. Memo: docs/research/CONDITIONS-AUDIT.md and the Q3 section of docs/research/anticipation/O1-CANDIDATES.md.

Decision. Three scored rows take role: 'condition' and are published beside their dimension under D122, unscored: vocational_secondary_share beside Learning, and labor_force_participation and electricity_transmission_losses beside Adaptability. Each keeps its World Bank route, is fetched on every ingest and is published with its value, year and rank among the countries that have it. D122 let a Tier B row move only on its own decision and only once its dimension had a capability row to take its place. This entry is that decision for the three, one paragraph each, and the replacements are already scored.

Vocational share of secondary enrolment (Learning). The share of secondary pupils on a vocational track describes how a school system is built, not what anyone learns in it, and the registry's own note said higher is not unambiguously better. It was the row that pulled Korea and Japan down in A7. Its latest year is 2019. Learning now reads the Human Capital Index, firm training and research citation impact (D124), which is the capability row D122 asked for. As a condition it correlates with log GDP per capita at 0.26 (n 51) and with the new Learning score at 0.28 (n 52).

Labour force participation (Adaptability). How much of the working-age population is in the labour force is a level set by norms, schooling and age structure. Its note called it "how much of the population can be reallocated at all", which is a description of what the economy has to move, not of it moving, and participation norms confounded it for India in particular. Adaptability reads reallocation through long-term unemployment (D120) and the breadth of the export basket (D119), beside the unemployment rate. As a condition it correlates with log GDP per capita at 0.575 (n 51) and with the new Adaptability score at 0.23 (n 53).

Electricity transmission losses (Adaptability). The state of the grid, and whether the operator can bill what it delivers, is infrastructure a country has to change with, which is the reason D122 gave for moving broadband. It now sits beside broadband in the same layer. It correlates with log GDP per capita at 0.69 (n 51) and with the new Adaptability score at 0.36 (n 53).

Why. Construct, under D118. Each row records what a country has, not what it does, and each dimension now has scored rows that observe the capability itself. The correlations below are printed as findings and decided nothing.

Measured on the 8.0.0 run against 7.8.0, same observations:

Learning's confidence rises although it loses a row, because the row it loses was observed at 2019 and the coverage denominator shrinks with it. Cuba falls to one observed Learning row and publishes no Learning score. Learning moves most at the top: Singapore 67.7 to 84.2 (seventh to first of 52), the United Arab Emirates 49.9 to 73.6, Canada 58.9 to 76.1 and the United States 45.5 to 60.7; Bolivia falls 57.2 to 35.9 (14th to 38th) and Honduras 38.2 to 33.2. Brazil moves from 28.0 to 35.0, 45th to 39th. Adaptability moves most where unemployment is low and exports are broad: India 59.5 to 85.4 (44th to tenth), Mexico 69.2 to 88.0 (32nd to sixth), El Salvador 69.4 to 86.5, Cuba 63.7 to 82.9 and Honduras 43.8 to 70.5; Switzerland falls 75.3 to 65.3 (20th to 43rd), the United Arab Emirates 74.7 to 68.7 and Nigeria 57.4 to 45.5. Brazil moves from 65.4 to 73.7, 36th to 35th. Across the release, with D142 and D143, the one-factor share (D137) falls from 0.523 to 0.485 and its correlation with income from 0.854 to 0.845; Adaptability's loading falls from 0.742 to 0.424. The guardrail, mean confidence across dimensions against log GDP per capita, moves from 0.278 to 0.289 (n 51). Every D138 reading is unchanged: structure, peers differ, stability mixed, leave-one-out robust, income explains part, the weak claim holds.

Cost. Adaptability's confidence falls by 0.12 and its correlation with income by 0.28, and the second is not a gain to claim: the rows that left carried income, and what remains is thinner. Nine countries read Adaptability on two rows, the unemployment rate and export concentration, because the ILOSTAT gate holds or lacks their long-term unemployment: South Korea, India, Mexico, Peru, Uruguay, China, the Philippines, El Salvador and Haiti. Four of them, Mexico, China, El Salvador and India, are in the top ten. The unemployment rate's own note calls it blunt, since low unemployment can mean a rigid labour market as easily as a fluid one, and on two rows it is half the score. The Brazil adaptability report reads Adaptability without the grid and the participation level in the score; both remain on the page as conditions. Published scores move in two dimensions, so this is a major version under D37 and 7.x numbers are not comparable.

Overturned by. A use measure showing that one of the three tracks a capability rather than a stock, such as grid losses moving with reform episodes the Adaptability rows also see, which would return that row to the score by its own decision; or evidence that the unemployment rate on two rows orders countries by labour-market rigidity rather than reallocation, which would make the two-row Adaptability scores a known artefact and argue for holding them until long-term unemployment covers them.

D142: Business share of R&D is retired: the only working source reads the make-up of a spending stock

Recorded 2026-10-02. Under D100 and D118. Issue #25. Dataset 8.0.0. Memo: docs/research/experimentation/O1-CANDIDATES.md, candidate 6.

Decision. business_rd_share (Experimentation) moves from gap to ingest: 'retired'. It stays in the registry and on every country's indicator list with status: 'retired', its note names the evidence, and it leaves the coverage denominator under D100. Its source field now names the dataset that was inspected, OECD MSTI with RICYT, in place of UNESCO UIS, which stopped publishing the series in March 2023.

Why. Construct first. Under D100 a gap says nobody publishes a comparable series and a retirement says a series exists, was inspected and measures the wrong thing. This row now fits the second. RICYT's open API serves R&D by sector of performance for Latin America, Brazil included, and with OECD MSTI the splice reaches 34 of 53 countries, 33 of them at 2019 or later. What it reads is the share of a country's R&D spending that business performs: the composition of a spending stock. D122 already moved R&D spending itself out of the scores as a condition, and the share of that stock is a description of the same thing. It is high where business research is large (Israel 94) and where total research is tiny and one firm dominates it (Thailand 80 and Vietnam 73 in the UIS archive). State enterprises count as business, so China and Vietnam read their ownership model, not firms experimenting. The splice's correlation with log GDP per capita, 0.742 (n 34), and with R&D intensity, about 0.75, are printed as findings; the retirement does not rest on them, and it does not rest on Experimentation's O1 shortfall.

Experimentation's denominator falls from nine rows to eight. Its mean confidence moves from 0.271 to 0.305 with no new observation, and no score moves. That rise is what D100 does by design, and it is not why the row is retired: AGENTS.md is explicit that gaps are not removed to make numbers look better. Experimentation still misses O1.

Cost. The benchmark no longer asks for a series it once named as the next best candidate for Experimentation in A1, and a reader who wants the corporate side of research has to read the R&D spending condition beside Anticipation. A version of the share restricted to private enterprises, if a publisher ever separates them, would be a different series and would need a new row.

Overturned by. A publisher that separates state from private enterprise in R&D by sector of performance across at least half the frame, which would remove the ownership confound and reopen the row as a gap; or an argument that the composition of R&D spending observes firms experimenting rather than what a country spends, which would make it a condition beside Anticipation's R&D spending instead.

D143: National belonging is retired: the only cross-national item reads pride and fails the regime test

Recorded 2026-10-02. Under D100 and D118. Issue #67. Dataset 8.0.0. Memos: docs/research/shared-purpose/O1-CANDIDATES.md, candidate 2, and docs/research/shared-purpose/EVS-WVS-BEHAVIOURAL-ITEMS.md.

Decision. national_belonging (Shared purpose) moves from gap to ingest: 'retired'. It stays in the registry and on every country's indicator list with status: 'retired', with the evidence in its note, and it leaves the coverage denominator under D100. Its source field names the item inspected, G006 in the Joint EVS/WVS.

Why. Construct first. The row asked for reported pride in and identification with the national community, and its own note said high national pride is not the capacity for collective action and must not be read as such. The Joint EVS/WVS sweep read G006, national pride, the only cross-national item aimed at it, and it fails A13: electoral autocracies read 75% very proud against 47% in liberal democracies, so scored it would lift the regimes that cultivate uniformity, which is the failure D121 kept polarization out of the score for. The other candidate found, ISSP 2023's closeness-to-country item, is also a perception and covers 16 countries. Under D100 that is a series inspected and rejected, not a measurement nobody can make.

Shared purpose's denominator falls from six rows to five. Its mean confidence moves from 0.343 to 0.411 with no new observation, which clears O1 on the arithmetic of D100 rather than on evidence. That effect is not the reason. The case rests on construct: the project does not believe pride is a capability a behaviour could one day observe under this name, and the observable acts of a shared project, civic participation (D128), tax revenue and, as a gap, volunteering, stay. No score moves.

Cost. Shared purpose now reads O1 as met while it rests on the same three rows it had at 7.8.0, and a reader comparing confidence across releases will see a rise with nothing behind it. A5 says so. The registry no longer names identification with a national community as part of the dimension, and a reader who thinks it belongs there has to argue it as a new row with a behavioural construct.

Overturned by. A cross-national item on belonging that is behavioural or that reads the same in every regime class, for example one that sorts liberal democracies and autocracies alike, which would reopen the row as a gap with that item as its candidate; or evidence that volunteering_rate and civic participation cannot carry the dimension without it.

D144: Customs clearance time is published beside Coordination as a behavioural check

Recorded 2026-10-02. Under D60 and D118. Issue #66. Dataset 8.0.0. Memo: docs/research/coordination/O1-CANDIDATES.md, recommendation 1 and its preflight.

Decision. World Bank IC.CUS.DURS.EX, the average days to clear direct exports through customs reported by manufacturing firms that export directly, in the Enterprise Surveys, is published beside Coordination as customs_clearance_time in checks.ts, lower_better, unit days, ingest: 'worldbank' from World Development Indicators (source 2). It is observed under __check__customs_clearance_time and enters no frame, mean, coverage count or confidence. The 8.0.0 ingest added 151 country-years and restated nothing. It covers 50 of 53 countries (no survey for the United Arab Emirates, Cuba or Haiti), latest 2025.

Why. On construct it passes: firms report the time the border took on their own shipments, which is the cooperation of agencies Coordination asks about and the current successor to the Doing Business border time frozen at 2019. It passes the A13 regime test as well (China, Vietnam and Singapore sit in the middle of the frame). It fails two other tests, and those are the reasons it is a check:

1. The sample base cannot be verified. The number of firms behind each country's mean, and its standard error, are published only in the microdata, behind a registration. On a proxy from the published tables, firms surveyed times the weighted share exporting directly, 37 of 53 countries reach 30 firms. The 12 held, ten of them in Latin America, are Panama, Nicaragua, Rwanda, Mexico, Honduras, Israel, Ecuador, Uruguay, the Dominican Republic, Bolivia, the Philippines and Paraguay. The proxy itself cannot settle the cases near 30, so the recommendation's own gate, 40 countries, is neither met nor refuted. 2. The questionnaire broke in 2024. Rounds fielded from 2024 add an all-agency release question, and since July 2025 the publisher drops every answer where customs time exceeds it. A country surveyed on both sides moves for reasons that are not its border: India 17.3 days in 2022 to 2.5 in 2025 on more than 9,000 firms each round, France 10.4 to 2.2. No benchmark country has two rounds on the new questionnaire, so stability within one design cannot be tested.

Its income correlation is printed and is not a reason: -0.12 on the published value against log GDP per capita (n 50), -0.28 against the Coordination score, from behaviouralChecks. No Coordination score, confidence or count moves.

Cost. D60's mechanism now holds a check kept out on sample and comparability grounds, alongside the income and construct reasons D60 and D121 used; the glossary entry says so. A published value outside the score will be quoted as a finding, including the twelve held countries and the cross-questionnaire moves, and the attached note is the only guard. The World Bank API lags the Enterprise Surveys portal, so the check can show an older round than the portal (Argentina 2017 against a 2026 portal round).

Overturned by. Reopened as an indicator question when either holds: someone with microdata access reads the per-country item base, and 40 or more countries pass at 30 firms on the real base; or a second round on the 2024 questionnaire lets stability be tested within one design. If it is reopened, prefer the all-agency measure, which asks about every border agency and needs a pinned adapter from the Enterprise Surveys portal. Retired instead if readers treat the check as a score, as D60 provides.

D145: New public repositories are scored in Experimentation from the GitHub Innovation Graph behind an access gate

Recorded 2026-10-02. Under D117 and D118, in the pattern of D120 and D124. Issue #65. Dataset 8.1.0. Memo: docs/research/experimentation/O1-CANDIDATES.md, candidate 1 and recommendation 1. D125 stays held: no GEM extension.

Decision. new_repositories_per_million joins Experimentation as a scored row: class O, higher_better, tier academic_survey, transform per_million_population over SP.POP.TOTL, unit "per million people". The value is the change in GitHub's count of public repositories located in the country from the latest first quarter a year earlier to the latest first quarter, read from data/repositories.csv in github/innovationgraph (CC0). The adapter github-innovation-graph-new-repos-v1 runs as pnpm bench github fetch, emits the count, and leaves the division by population to the scorer, as the patent, trademark and design rows do. The publisher restates its files in place every quarter and keeps old versions only in git, so the pin is the commit: GITHUB_IG_COMMIT in model/source-catalog.ts, here 054c7dbc ("release q1 2026 data", 2026-07-07), which /sources prints as the row's link. The observation file carries the pin under github: repository, full SHA, path, retrieval date, latest quarter in the file, the window, the gate's parameters, and every benchmark country's first-quarter stock since 2020, so a rescore never touches the network. --commit latest names the newest commit; moving the pin is an edit to the constant and a revision run.

Before emitting, an access gate runs on every country and names none. A country whose stock grew by less than a quarter of the median benchmark country's growth over the same window is held: it gets no value, never a zero, and the file records its growth, the threshold and its year-on-year changes under held. On the 2026 Q1 file the median is 24.6% and the threshold 6.2%. China (down 0.1%, the stock fell in four of six yearly windows) and Cuba (up 2.8%) are held; the next lowest is the United Kingdom at 12.9%. 51 of 53 countries are scored on the row. Brazil adds 3,925,512 repositories, 18,446 per million, 17th of 51, normalized 25.5. Glossary: "Access gate".

Why. Construct first, written in the memo before values were read in. Experimentation asks whether a country makes many cheap attempts and abandons them freely. A public repository costs nothing to start, needs no fee or lawyer, and is kept whether or not it is abandoned; a yearly count per head is the closest open measure of that, of the same kind as trademarks and designs per million. It is behaviour, the change in a year; the level, developers or repositories per million, is platform adoption and was rejected in the memo. It covers every benchmark country at a current date, and its rank order is stable year to year (Spearman 0.97 to 0.98).

The gate is a rule because the exclusion has to be auditable, as D120's is. The stock is net of deletions and relocations, so a stock that stalls while the platform grows by at least 12.9% everywhere else records where developers can or do host code, not how many projects they start: China's developers work behind network filtering and on domestic platforms, and Cuba's access to the platform has been constrained by US trade controls. GitHub's own trade-controls page now lists its services as generally available in Cuba, so a sanctions citation alone would not hold the row out; the stalled stock is the evidence, and a rule on it applies to any country that stalls next.

Reported as findings, not tests (D118), on dataset 8.1.0 against 8.0.0. The row's normalized value correlates with log GDP per capita at r = 0.701 (log10 of the per-million value 0.815, n 50), and its wealth-attribution delta is 0.080. Experimentation's correlation with log GDP moves from 0.572 to 0.651 (Spearman 0.726 to 0.801, n 51), a finding against the claim in this dimension. Its mean confidence moves from 0.305 to 0.364, still under O1, and the bottom quarter from 0.192 to 0.253; 53 countries are scored either way. The guardrail, mean confidence against log GDP per capita, moves from 0.289 to 0.286 (n 51). The shared factor's share moves from 0.485 to 0.498 and its correlation with income from 0.845 to 0.847. Under D138 no reading changes: the residuals still share structure (0.287 against a 95th of 0.217), peers still differ, release order stays mixed, income is still part of the distance and the weaker claim holds; Experimentation's residual loading rises from 0.05 to 0.17. No redundant pair forms. No other dimension's score or confidence moves.

Cost. One platform and public work only: GitLab, Bitbucket, Gitee and private repositories are invisible, and the publisher is a firm reporting on its own platform, for which no source tier fits; academic_survey is the nearest. Location is the members' modal IP location, so VPN users are misplaced and Singapore's 260,665 per million is partly a hub; the upper Tukey fence caps it at 100, with Estonia next at 89.8. Coursework, bootcamps and tutorials create repositories and may lift countries with large student cohorts. The quarter-of-median threshold is judgment, and the gate reads only the latest window, so a country whose count resumes growing is scored as soon as one window clears. Held countries carry an unmeasured scored row: China's and Cuba's Experimentation confidence falls from 0.267 to 0.237, and their scores rest on the other rows (China 100, Cuba 1.1, unchanged). The largest score moves are Ireland +14.9, Germany -13.5, Turkey -13.3, the Netherlands +12.1 and France -12.1; Brazil's score does not move (25.6). The value is stamped with the closing first quarter's year and divided by the latest population, which is one year older.

Overturned by. Evidence that new repositories track platform adoption rather than attempts, for example a within-country series in which the count follows developer sign-ups and not new projects, which would make the row a behavioural check under D60. A second code-hosting source with frame coverage that reorders the frame, which would show the row reads one platform's share of the market. A held country whose count resumes growing past the threshold for a reason that is not access, or an unheld country falling under it for a reason that is about capability, either of which would show the gate removes real values. Or GitHub changing what repositories counts or how it locates one, which needs a new adapter version and a new entry.

D146: The foresight register is piloted before it is coded, and the pilot sends the codebook back for revision

Recorded 2026-10-02. Follows #68. Codebook 1.0 is 2cc0577, committed before any coding.

Decision. government_foresight_capacity stays a declared gap. The project-coded register of government foresight functions was piloted on ten countries (HTI, VNM, ARE, NIC, SGP, NGA, BRA, GBR, EST, CHE), drawn across the four V-Dem regime classes and the income quartiles with a fixed seed and Brazil fixed, by two independent coders. Codebook 1.1 must settle the eight ambiguities the pilot found, then the same ten are recoded, before any decision to code the 53 or to score the row. Pilot: docs/research/anticipation/FORESIGHT-REGISTER-PILOT.md and data/research/foresight-register-pilot.json.

Why. I1, established by a named act, reached Krippendorff alpha 0.79 against the 0.80 threshold fixed before coding (0.69 on the six countries with a function). I4 never fired. Brazil's agreed score of 100 hid a disagreement over which bodies qualify. The survival rule, any change of the person holding executive power, did not reward closed regimes: closed autocracies average 11 and liberal democracies 100, regime rho 0.72, under the 0.80 stop rule. The named-act rule leans towards states that legislate their foresight bodies, which scores Singapore 67 on documentation rather than function. Income r 0.69 (n 10) is printed and decided nothing.

Cost. A second coding round before any row exists, so Anticipation keeps two scored rows that track income (r 0.872) at least that long. Both coders were one model, so their errors are correlated, and even a passing recode needs a human coder, or a 20 percent human second coding on the frame, before promotion.

Overturned by. A recode under 1.1, with at least one human coder, reaching alpha of at least 0.80 on every item and on body-by-body decisions, with no ceiling and regime rho under 0.80, which supports coding the 53 at tier expert_panel. Regime rho at or above 0.80, or eight of ten at one score after revision, stops the register.

D147: Two evidence records leave the corpus with the row they were filed against

Recorded 2026-10-02. Follows D142, which retired business_rd_share.

Decision. gbr-business-rd-share and prt-sifide are deleted from data/evidence/records.json. Both were filed against business_rd_share, and no declared gap fits them: the first is the retired statistic itself, the second an R&D tax incentive. They remain in git. bra-embrapii and bra-sibratec are refiled under university_industry_collaboration, whose construct they fit. bench validate now warns on any record filed against a retired row, which is how these went unnoticed for one release.

Why. A retired row is a measurement the project declined to make (D100), so a record filed against it stands against nothing. EVIDENCE.md requires deleted ids to be named in a decision entry so the corpus history stays auditable.

Cost. Two delivered cases leave the published corpus. Neither was a reversal, so the quota is unaffected.

Overturned by. business_rd_share returning as a gap or a scored row, which would make both records admissible again.

D148: The foresight register goes to the 53 under agent coding and a human spot-check

Recorded 2026-10-02. Follows D146 and #68. Codebook 1.1 is 92ff79e, committed before recoding; round 2 is c371654.

Decision. Coding of government_foresight_capacity proceeds to the 53 countries once a person has checked the sample in docs/research/anticipation/FORESIGHT-REGISTER-SPOTCHECK.md and it passes. The row stays a declared gap and is not scored until a further decision promotes it at tier expert_panel. Before the 53 are coded, codebook 1.2 settles whether a parent body is coded separately when its foresight unit outlives it. This supersedes D146's requirement of at least one human coder: two agent coders of different model families code, and a person checks a seeded, regime-stratified 20 percent of bodies before anything is scored (owner ruling, 2026-10-02).

Why. Under codebook 1.1, which applies the owner's ruling that I1 accepts any official record dating a body's creation, two fresh coders agreed on every item in all ten pilot countries: alpha 1.00 on I1 to I3 and on the score. I4 is 0 throughout, a floor. Body-by-body alpha is 0.83 on the measure fixed in 1.1, and the three split bodies are closed bodies that change no item. Regime rho fell from 0.72 to 0.64, under the 0.80 stop rule, and no ceiling fired (five at 100, four at 0). Income r 0.83 (n 10) is printed and decides nothing.

Cost. Both coders are one vendor's models, so their errors stay correlated, and a 20 percent spot-check is weaker than a human second coder. Body alpha is 0.79 on the bodies both coders judged, under the threshold on round 1's measure. I4 has never fired, so its reliability is unmeasured. On the ten, I1 and I2 move together and the score is close to binary: if the 53 confirm that, the row reads whether a qualifying body exists, with survival on top, and at r 0.83 on the pilot it will not lower Anticipation's income correlation.

Overturned by. A spot-check that overturns two or more of the 11 sampled bodies, or any body decision that changes a country item. Or, on the 53: alpha under 0.80 on any item or on body-by-body decisions in a 20 percent second coding, regime rho at or above 0.80, or 43 or more of the 53 at one score.

D149: New export products are scored in Adaptability from the Growth Lab's Atlas trade data

Recorded 2026-10-02. Under D117 and D118, in the pattern of D119 and D124. Fixes A16 in part. Memo: docs/research/adaptability/A16-FIXES.md, recommendation 1 and the preflight of 2026-10-02. Dataset 8.2.0.

Decision. new_export_products_rate joins Adaptability as a scored row: class O, higher_better, tier academic_survey, unit "% of products not exported competitively at the start". A product is new when its revealed comparative advantage averaged under 0.5 over 2009-2011 and at least 1 over 2022-2024, with mean exports of at least USD 1 million a year over 2022-2024. The rate is new products over the products under 0.5 at the start, stamped 2024. The source is the Harvard Growth Lab's Atlas of Economic Complexity, International Trade Data (HS, 92), doi:10.7910/DVN/T4CHWJ version 18.0 (2026-04-22, CC0), file hs92_country_product_year_4.csv. The adapter atlas-hs92-new-export-products-v1 runs as pnpm bench atlas fetch, downloads the file by its Dataverse id, refuses it unless its MD5 and SHA-256 match the pin in model/source-catalog.ts, and reads only the two windows. Only four-digit codes are products, so the unspecified XXXX line is never one. The observation file carries the pin under atlas: release, checksums, the rule, the product universe (1,242) and, for every country, the products it had room to enter, the new ones by code, and how many of them it exports more of than it imports. Glossary: "New export product".

Preflight. The owner made wiring conditional on two traps. Entrepot hubs: the file has gross exports only, so the test was a net exporter variant (a new product counts only where exports exceed imports). It orders the frame at Spearman 0.93 against the published rate; the hubs keep 0.56 to 0.69 of their new products against a frame median of 0.75, and so do open economies inside value chains (Poland 0.60, Portugal 0.50, El Salvador 0.40); no hub's rank moves more than the frame's median move of four places. The row is scored on gross exports and the net count travels in the pin. Sanctioned trade: Cuba and Venezuela, and Haiti, which is not sanctioned, did not report to UN Comtrade in one or both windows and are read from partner reports. A sanctions regime is a shock an economy reallocates under, so their values are a reading, not an artefact; a gate keyed to sanctions would need a list and one keyed to reporting would catch Haiti. 19 of Venezuela's 21 new products grew in absolute value; across the frame 6 of 1,381 crossed the line only because a basket shrank. Both are scored, with no gate.

Why. Construct first, written before values were read. Adaptability asks whether an economy can absorb a shock and reallocate, and entry into new export lines is reallocation observed: capital and workers moving into products the country was not making. It is behaviour, not a stock money buys, and counting over the room left does not reward a basket that was already broad. It covers all 53 countries at 2024 from one release, its order is stable across windows (Spearman 0.92 to 0.94), and it is the only candidate in the memo that gives the nine countries the ILOSTAT gate holds a third row on evidence about reallocation rather than slack.

Reported as findings, not tests, on 8.2.0 against 8.1.0. The row's normalised value correlates with log GDP per capita at 0.18 (n 51); its wealth-attribution delta is -0.025. No redundant pair forms; its overlap with export concentration is 0.54 normalised.

The nine of A16 move from two rows to three and from confidence 0.380 to 0.459: India 85.4 (10th) to 83.3 (4th), China 87.0 (7th) to 80.2 (7th), South Korea 83.4 (13th) to 72.7 (12th), the Philippines 74.5 (34th) to 69.9 (22nd), Mexico 88.0 (6th) to 64.3 (28th), El Salvador 86.5 (8th) to 62.6 (32nd), Peru 70.4 (39th) to 50.7 (46th), Uruguay 64.8 (44th) to 46.3 (49th), Haiti 31.6 (52nd) to 22.5 (53rd). Every Adaptability score moves, because a quarter of each is now the new row (rank Spearman 0.80 between releases): the United States 88.4 (5th) to 71.6 (17th), Cuba 82.9 to 62.2, Ecuador 80.9 to 62.4, the United Arab Emirates 68.7 (41st) to 72.0 (16th); Brazil 73.7 (35th) to 61.4 (37th). The top ten becomes Poland, Vietnam, Thailand, India, the Netherlands, Israel, China, Turkey, Estonia and Malaysia. No other dimension's score or confidence moves.

The guardrail, mean confidence across dimensions against log GDP per capita, moves from 0.286 to 0.274 (n 51). The shared factor (D137) carries 0.496 of the variance, from 0.498, and correlates 0.845 with income, from 0.847; Adaptability's loading moves from 0.424 to 0.400. Under D138 no reading changes: the residuals still share structure (0.286 against a 95th of 0.217), peers still differ, release order stays mixed, leave-one-out is robust, income is part of the distance and the weaker claim holds; Adaptability's residual loading moves from -0.284 to -0.255.

D119's overturning clause said a source observing entry into new export products would replace export concentration. This entry keeps both, and supersedes that half of the clause: concentration reads the exposure a single price shock meets, entry reads the move away from it, and the two overlap at 0.54, well under a duplicate.

Cost. Merchandise only, so services exporters (the United States, the United Kingdom, India's software) are read on goods. Exports are gross, so a hub counts what passes through it, inside the noise the preflight measured but not zero. A resource find counts as entry when it crosses the line. The thresholds (0.5, 1, USD 1 million, three-year means, fifteen years) are judgment fixed in the memo before wiring. The window is slow, so the row moves little between releases and has no trend. The publisher is an academic group, for which academic_survey is the nearest tier. Three countries rest on partner reports. The fetch downloads 452 MB. A16 is restated, not closed: the nine still have no long-term share.

Overturned by. A domestic-exports series (re-exports removed) at frame coverage that reorders the hubs by more than the frame's typical move, which would make the gross rate a hub artefact and call for that series or a gate. Evidence that entry counts track resource finds or a shrinking basket more than reallocation, for instance a window in which most new products are minerals or crossed the line on a falling total. Or a new release of the file that changes the definition of RCA or the reconciliation, which needs a new pin and a revision run.

D150: Informal employment is published beside Adaptability as a condition

Recorded 2026-10-02. Extends D122 and D141. Memo: docs/research/adaptability/A16-FIXES.md, recommendation 2. Dataset 8.2.0.

Decision. informal_employment_share joins the registry with role: 'condition' beside Adaptability: informal employment as a share of total employment, ILOSTAT SDG indicator 8.3.1 (DF_SDG_0831_SEX_ECO_RT), both sexes, whole economy, the latest year since 2010, a labour force survey preferred where ILOSTAT holds more than one for a year. It is read by the ILOSTAT adapter through the SDMX endpoint D120 uses, as pnpm bench ilostat fetch --only informality, into data/observations/ilostat-informality.json, and published with its value, year and rank among the countries that have it, never on the 0 to 100 scale. Direction is lower_better, which orders the rank and claims nothing about Adaptability. The World Bank API does not carry the series at comparable coverage: pnpm bench probe finds non-agricultural informality in the Gender Statistics database at 24 of 53 and the Jobs database's own measure at 29 of 53, latest 2021.

Why. Construct. The share of work with no contract, registration or social protection describes the labour market a country reallocates through, not the reallocation. It is the context A16 asks for beside a low unemployment rate: India 87.2%, Peru 70.5%, El Salvador 63.6% and Mexico 56.9%. It fails the indicator test on direction: informal work is both the slack that absorbs a shock and the precarity that keeps a worker from moving up. That is what D122 makes a condition, as D141 did for labour force participation. It changes no score or confidence. As a finding, it correlates with log GDP per capita at 0.90 in its ranked direction (n 42) and with the Adaptability score at 0.30 (n 43).

Cost. 43 of 53. ILOSTAT has no value for the United States, Canada, Australia, Japan, Singapore, Israel, Malaysia, China, the Philippines and Cuba, two of them among the nine A16 names. Some values are old (Haiti and Nicaragua 2012, South Korea 2019). Comparability is weak: European values come from EU-SILC, an income survey with a proxy definition, which the observation note says. ILOSTAT revises in place, so the retrieval date is the release identifier. In this frame the condition is very nearly income.

Overturned by. A defensible direction, for instance evidence that informality reliably slows or speeds reallocation after a shock across countries, which would make it an indicator on its own entry. Or a harmonised series that closes the EU-SILC gap and the ten missing countries, which would make the rank comparable enough to read across income groups.

D151: A construct still wanted stays a gap when its only series fails a construct test; national belonging is reopened

Recorded 2026-10-02. Amends D100. Supersedes D143. Leaves D121 and D132 standing. Owner decision of 2026-10-02. Memo: docs/research/GAP-AUDIT.md, classes (b) and the notes on D121, D132 and D143. Dataset 8.3.0.

Decision. D100's description of a retirement gains a rule. A row is retired when the project no longer wants its construct measured, or when the row's definition is not a capability: a row defined as one rejected dataset (the WGI and LPI rows of D23), or as the make-up of a spending stock (business_rd_share, D142). A construct the benchmark still wants stays a declared gap when the only series aimed at it is inspected and fails A13 or another construct test. The failed series is recorded in the row's note and, where it is worth showing, published as a check under D60, and the row stays in the coverage denominator.

Under that rule institutional_trust (D132) and political_polarization (D121) stay declared gaps, unchanged, with their checks beside them. And national_belonging moves from ingest: 'retired' back to ingest: 'gap'. Its definition drops pride: it now reads reported identification with and attachment to the national community, as distinct from pride in the nation. Its note names G006 as the inspected item that is not wired and ISSP 2023 as the 16-country lead. The rows retired under D23, D44 and D142 are not reopened by this entry; a challenge to any of them is a new decision.

Why. The audit found three rows tested the same way, each with one cross-national series that reads the definition and fails A13 because closed or electoral autocracies read best. D143 retired its row and D121 and D132 kept theirs, and D100 as written sided with D143. The owner's position is the one D121 and D132 took: the benchmark wants as many relevant constructs read as it can get, and a construct is not abandoned because the first instrument aimed at it flatters uniformity. D143 gave a reason of its own, that pride is not a capability. That holds for pride, and the reopened definition no longer asks for it. Whether people count themselves members of the community whose common project Shared purpose asks about is still wanted, and nobody can measure it comparably yet, which is what a gap says.

The rule also closes the hole D100's overturn clause warned about. A retirement now records a judgement on the construct, not on the first proxy tried. Without the rule, any construct could leave the denominator by trying one bad series, and the confidence of the dimensions that most need measurement would rise for the attempt.

Cost. Shared purpose's denominator rises from five rows to six. Its mean confidence falls from 0.411 to 0.343, under O1's 0.40 again, with no observation lost. That is the honest state: Shared purpose rests on three scored rows, the same three it had at 7.8.0, and the confidence 8.0.0 to 8.2.0 published came from a retirement and not from evidence (A5 said so). 28 countries' Shared purpose confidence falls from usable to thin and two from thin to very thin, so their agendas move the dimension from raise to measure; China's leading item changes from Shared purpose to Anticipation, and Switzerland and South Korea are left with no raise item. No score moves. The guardrail moves from 0.274 to 0.272. Three dimensions now miss O1: Coordination, Experimentation and Shared purpose. The bump is minor, because a registry row changes status and no country or published field changes, as D100's own restatement was.

Overturned by. Evidence that readers take a declared gap with a failed series beside it as a measurement the project made, which would argue for a third status between gap and retired; or an owner decision that a specific construct is no longer wanted, which retires that row under this rule. For national_belonging: an item on belonging that reads the same in every regime class, or a behaviour of belonging, for 27 or more of the frame, would make it an indicator candidate.

D152: Four gaps that read a stock are redefined as behaviour and stay declared gaps

Recorded 2026-10-02. Extends D118 and D122; under D151. Owner decision of 2026-10-02. Memo: docs/research/GAP-AUDIT.md, class (c). Dataset 8.3.0.

Decision. The four gaps the construct audit classed as not a capability keep their ids and their dimensions and stay ingest: 'gap', under new definitions that name an act a series could be tested against.

• basic_research_share (Anticipation), now "Long-horizon research commitments": new competitively awarded public research grants that run five years or longer, as a share of all new competitive public grants awarded in the year. Unit % of new grants, class C. It was the share of R&D spending classed as basic research, the make-up of a spending stock that D142 retired business_rd_share for. No grant register has been triaged.
• venture_capital_gdp (Experimentation), now "Venture deals": first venture capital rounds closed by companies based in the country in the year, per million people, counted by deal and never by amount. Unit per million people, class C. It was venture capital deployed as a share of GDP, a level of financial depth of the kind domestic_credit_private is (D122). Candidates: the OECD SME and Entrepreneurship Financing scoreboard (6 of the 16 countries checked under D21, amounts not deals), national association counts on their own definitions, and commercial deal databases that fail D10.
• adult_digital_skills (Agency), now "Adults doing digital tasks": the share of adults who carried out a named digital task in the last three months, counted one task at a time and never as a skill level. Unit % of adults, class C. It was the share who can perform standard tasks, a skill level of the kind D122 moved out with internet_users. Candidate: ITU ICT skills by type (SDG 4.4.1) through the UIS API, probed 2026-10-02 at 33 of 53 for attaching a file, 25 at 2018 or later.
• regulatory_sandbox_activity (Experimentation), now "Firms through regulatory sandboxes": firms admitted to, and firms completing, a regulatory sandbox or controlled trial regime in the year, per million people, summed across regulators. Unit firms per million people, class C. It was the number and breadth of live sandboxes, a policy stock. Candidate: the cohort lists regulators publish one by one; the evidence corpus holds 22 records against the row, most of them admission or exit counts on each regulator's own definition. It stays an evidence grid column (D135).

Each note names the bar a series must clear: one definition, 27 or more of the frame. The es and pt-BR lexicons carry the new names and definitions, and two new units.

Why. D151 keeps a construct in the denominator while the project wants it. A row whose definition reads a stock is not a construct the project wants scored (D122 moved ten such rows out), so the audit's choice was between a condition, a retirement and a redefinition. The owner chose the redefinition: each row points at a capability the dimension lacks, and the stock wording was the first available proxy written into the definition, not the thing wanted. A first round counts an attempt with outside money where an amount counts the money; a grant length is a funder deciding to wait; a task done is an act where a skill is a level; a firm through a sandbox is a trial run where a regime is a permission. Each new definition is narrow enough that a candidate series can be held against it and fail.

The ids stay. venture_capital_gdp no longer describes its row, but evidence records, /indicators#venture_capital_gdp, data/out/indicators/ and any pinned consumer read the id, and renaming it would remove a published id, which D37 makes a major version. Its note says the id keeps its first name.

Cost. No confidence moves: each row was a gap and stays one, so the denominators of Anticipation, Agency and Experimentation are unchanged, as are all scores. The audit's alternative, retiring or moving the four rows, would have raised those three dimensions' confidence with no observation (the audit printed Anticipation 0.451 to 0.601, Agency 0.483 to 0.604 and Experimentation 0.364 to 0.468), and this decision declines that. Rows now ask for series that mostly do not exist: the grant and sandbox rows are registers Envisioning would have to build, and the venture row has no inspectable full-frame source. A1's fix now asks for a deal count rather than an amount.

Overturned by. A source that reads the old definition at full frame and an argument that the stock, not the act, is the capability; evidence that a redefined row's candidate series reads income and nothing else once wired, which would make it a condition under D122; or an owner decision that a construct is no longer wanted, which retires the row under D151.

D153: The frame takes every country of a million people that is measurable in all nine dimensions

Recorded 2026-10-02. Dataset 9.0.0. Supersedes D99's closing of the country list at 53 and the per-country reason that D27, D51 and D99 wrote for each addition. Extends D51's rule-based completion to the world. Re-baselines the guardrail beside O1 in D117. Applies D47: a major version.

Choice. The benchmark holds 125 countries: the 53 it held at 8.3.0 and every World Bank economy of at least one million people (SP.POP.TOTL, latest year) that reaches MIN_INDICATORS_FOR_SCORE observed rows in all nine dimensions on the 8.3.0 registry with the adapters already wired. That rule picked 72 countries in docs/research/FRAME-EXPANSION.md (set C), and the rescore confirms it: all 72 publish nine scores at 9.0.0. In population order they are Pakistan, Bangladesh, Russia, Egypt, the Democratic Republic of the Congo, Iran, Tanzania, Italy, Myanmar, Sudan, Uganda, Iraq, Angola, Ukraine, Morocco, Uzbekistan, Mozambique, Ghana, Madagascar, Côte d'Ivoire, Nepal, Mali, Burkina Faso, Malawi, Zambia, Sri Lanka, Kazakhstan, Romania, Zimbabwe, Guinea, Burundi, Tunisia, Belgium, Jordan, Czechia, Tajikistan, Papua New Guinea, Greece, Azerbaijan, Hungary, Austria, Belarus, Laos, Kyrgyzstan, Serbia, the Republic of the Congo, Bulgaria, Denmark, Lebanon, Norway, Slovakia, New Zealand, Georgia, Croatia, Mongolia, Bosnia and Herzegovina, Namibia, Armenia, Lithuania, Jamaica, the Gambia, Botswana, Lesotho, Moldova, Albania, Guinea-Bissau, Slovenia, Latvia, North Macedonia, Cyprus, Trinidad and Tobago and Mauritius.

Out by the same rule: the 22 economies that clear eight dimensions (the seven OAPI members with no national IP office series, the Gulf monarchies with no Gini or tax ratio, and others), the 13 that clear seven or fewer, the 15 economies under a million that clear all nine (Iceland, Luxembourg, Malta, Montenegro and eleven others, most of them small islands), and Taiwan, which the World Bank API does not carry. The million line is the smallest country already in the frame (Estonia), the Growth Lab's own ranking floor, and the point below which each microstate would move the Tukey fences as much as India does; admitting them is a separate question with its own entry.

At 125 the reason field cannot be a case per country, and a case chosen per country is the selection question D51 removed for Latin America. Every one of the 72 carries one shared sentence that states the rule, and the original 53 keep theirs.

What changed in the code. countries.ts gains the 72 entries. UNCTAD_M49 gains their M49 codes (all 125 resolve). The Joint EVS/WVS adapter gains two source aliases: the release prints Uzbekistan as "Uzbequistan" and cuts Bosnia and Herzegovina's label at the column edge to "Bosnia and"; the adapter test that once reported that cut label as unmapped now holds it to the alias. The Spanish lexicon names every country, as its test requires, and the Portuguese one names the 72 too. COUNTRY_ROW_FACTS gains the ILOSTAT facts the pinned file states for the new countries: the Democratic Republic of the Congo and Botswana read from household surveys (Botswana's flagged unreliable), and Belgium, Croatia, Jamaica, Myanmar and Uganda carry an unreliable flag. The scoring test's permutation fixture assumed a prime country count; 125 is not, so it now checks that each row's multiplier is coprime with the count.

The ingest. Each source ran through its own command, so revisions.json logs every addition: World Bank 55,557 values added and 0 restated, Joint EVS/WVS 174, V-Dem 288, UNCTAD 72, ILOSTAT long-term unemployment 66 and informal employment 59, Growth Lab Atlas 72, GitHub 71. OpenAlex added 72 and restated 50 of the 53's values, by a mean of 0.010 on a ratio near 1 and at most 0.088 (Thailand), because its counts move between reads and the API has no versions (D124). The gates ran on the new frame:

• ILOSTAT's plausibility gate (D120) holds the same six countries it held before (South Korea, Mexico, Peru, the Philippines, El Salvador, Uruguay), and drops single years for seven new ones without emptying their rows: Bulgaria 2025 and Iran 2018, Madagascar 2022 and Mali 2024 as unconfirmed jumps, Egypt 2011, Georgia 2019 and Mauritius 2022 as spikes. Six new countries have no ILOSTAT value at all: Uzbekistan, Guinea, Tajikistan, Papua New Guinea, the Republic of the Congo, Trinidad and Tobago.
• The GitHub access gate (D145) takes its median over the benchmark, so it moved: median stock growth 26.9%, gate 6.7%. It holds China (minus 0.1%), Cuba (2.8%) and, new, Belarus (6.2%). Iran is not held.
• The D64 hold on countries with separate EVS and WVS rows now holds ten: Armenia, Czechia, Romania, Russia, Serbia, Slovakia and Ukraine join Germany, the United Kingdom and the Netherlands. The hold stays. The survey memo shows that pooling them would raise the guardrail to about 0.60, which is a finding for the pooling question, not a reason in it.

The statistics, at the D137 and D138 precision. At 8.3.0 the factor test ran on 51 complete cases and the residual tests on 50. At 9.0.0 they run on 123 and 122 (Cuba and Haiti stay out of the complete cases; Venezuela and Cuba have no income figure).

The interval on r is now about plus or minus 0.06, close to the 0.05 the memo predicted, and it still contains 0.8: the factor still reads as income, at "strong" by a margin smaller than its own uncertainty. D137's overturn clause asks for the bands to give way to an interval once r can be estimated within a few hundredths. This entry publishes the interval; the change to readFactorTest and the surfaces that print the band is a new reading rule and gets its own entry. The (b) margin narrowed from 0.043 to 0.036, but its null threshold fell from 0.233 to 0.190 as the memo said a wider set would, so the reading rests on a tighter estimate: the memo's resampling put its standard error near 0.02 at this size, against 0.03 to 0.035 at 50. (d) moves from "part" to "little": on the wider frame income accounts for under a quarter of the average country's distance from the average profile, though the pooled share is 0.411 against 0.399. (c) compares consecutive releases on the same country set, so 9.0.0 pairs with nothing and its history starts again with the next release on 125.

Correlation of each dimension with log GDP per capita, 8.3.0 (n 51) to 9.0.0 (n 123): Anticipation 0.872 to 0.863, Agency 0.579 to 0.466, Coordination 0.558 to 0.533, Trust 0.606 to 0.354, Learning 0.779 to 0.709, Experimentation 0.651 to 0.728, Adaptability 0.431 to 0.533, Building 0.434 to 0.573, Shared purpose 0.202 to 0.259. No dimension pair passes the redundancy threshold. 0 of 3,721 observed current cells clamp.

The guardrail is re-baselined at 0.526. Mean confidence across the nine dimensions correlates with log GDP per capita at r = 0.526 across the 123 countries with an income figure (slope 0.046 confidence per tenfold income), against 0.272 across 51 at 8.3.0. The owner re-baselines rather than refusing the frame, for this reason: 0.27 was partly a product of which 53 were chosen. The original set reached the poor countries the Joint EVS/WVS release happened to survey (Ethiopia, Nigeria, Kenya, Bolivia, Nicaragua, Guatemala), so survey presence correlated 0.04 with income inside it and 0.37 in the wider set; across all 154 measurable countries of a million or more the figure is 0.41. Confidence does not depend on the frame, and the 53 read 0.272 inside the 9.0.0 output too: the rise is who is measured, not how. The survey memo (docs/research/trust/SURVEY-COVERAGE-FOR-EXPANSION.md) decomposes it: with no survey items anywhere the figure would be 0.460, so the survey gap explains about 0.07 of it, and the rest comes from the IP office rows, the ILOSTAT row and the other thin columns. The pending repair is Afrobarometer Round 8, whose trust item is the Joint release's word for word and would take the guardrail to about 0.509 and Trust back above 0.40; it waits on a written licence confirmation. From here the guardrail watches drift from 0.526 inside the 125, which is what it was for: a source that only reaches rich countries raises it, and that rise is the warning.

O1. Mean confidence, 8.3.0 to 9.0.0: Anticipation 0.451 to 0.453, Agency 0.483 to 0.470, Coordination 0.362 to 0.371, Trust 0.418 to 0.398, Learning 0.513 to 0.523, Experimentation 0.364 to 0.318, Adaptability 0.577 to 0.580, Building 0.551 to 0.548, Shared purpose 0.343 to 0.319; across all nine 0.451 to 0.442. Trust drops under 0.40 by construction of the frame, because half its rows are survey items the new poorer countries lack, and joins Coordination, Experimentation and Shared purpose below the target. Experimentation falls furthest, because GEM stays at 16 (D125) and the national IP office rows thin out. The roadmap queue reads these figures from now on.

Evidence grid and foresight register. D135 called the corpus complete when every grid cell is closed. The grid is now 1,000 cells, 424 of them closed and 576 open, all 576 in the new countries; the corpus is not complete again until they close, and the research inventory prints them as the queue. The foresight register (D148) was planned as nine batches for the 53; the 72 add about twelve more under the same runbook. Neither backlog blocks the release: an open cell lowers no score and no confidence.

Delphi. No panel run is on this dataset. D139 already kept the 8.x session estimates off every comparison with 9.0.0 indicators. A full run over 125 countries prices at about $58 with the current four panelists.

The front-page field. FlagField now draws up to 125 flags on one axis. Checked at this release on the front page and a capability page, at desktop and phone widths: the stacks fit the frame and no flag overlaps another, so the component is unchanged. On a phone each flag shrinks to about eight pixels, which reads as a distribution but is a small target to tap; a fix stays inside the D67 component, never a second distribution.

Cost. Every published number from 8.3.0 is restated and is not comparable with 9.0.0. Of the 470 dimension cells scored in both, 462 moved, by a mean of 3.2 points; the largest moves are Experimentation for Israel (28.3 to 48.0), Sweden and the United Kingdom, which rise because the countries added below them widen the bottom of the IP office rows. Brazil moves on all nine dimensions, and its ranks restate against 125. The observation file grows from 17 to 40 MB, data/out from 20 to 42 MB, and index.json, which every list page reads whole, grows with it. The 72 arrive with no evidence records, no foresight coding and no panel estimate. The velocity and leverage sandbox fixtures were regenerated for the 125, the first time since 6.1.1, and on today's registry velocity finds a complete five-year series in at most four dimensions for any country, so it excludes all of them from its country-level read; that is the registry since 6.1.1, not the rebase.

Overturned by. A rescore on a later release whose D137 interval on r or D138 (b) margin is no tighter than this one, which would mean the added countries are noisier than the 50 they joined and the precision argument failed; a source that measures Experimentation for the OAPI members or Shared purpose for the Gulf, which would bring the eight-dimension group under the rule and reopen it; or a guardrail that rises above 0.526 inside the 125 without a country being added, which is the drift the re-baselined figure exists to catch.

D154: Open data and model knowledge only: no gated sources, and an in-session multi-vendor panel is a panel

Recorded 2026-10-03. Owner ruling of 2026-10-03. Extends D10. Supersedes the gated reopening routes in D64 and D144, the licence-confirmation step D153 named for the guardrail repair, the rule in docs/PANEL.md that a gateway run must replace an in-session run before publication, the four-vendor requirement for in-session runs (D13), and D148's human spot-check. Closes #73, #66 and #38.

Choice. The project works from two inputs: open data, and what language models know about the world. It makes no registrations, licence requests, accounts or emails to data owners.

(a) Open data only. A source enters the benchmark only if its data and the values the project derives from it can be downloaded without an account and republished under the project's licence (CC BY 4.0 for data/out) on the publisher's terms as published. A term that is silent on redistribution fails the test: the project does not write to ask. This extends D10, which asked whether a source can be inspected, with a second question, whether it can be taken and republished without asking anyone. Every source wired at dataset 9.0.0 was admitted with a licence note in its memo; one found to fail the test is reviewed under this entry in its own change, never kept silently. This closes:

• Afrobarometer Round 8 Q83 as a fill-only second source for A165 (#73). The data are free and under copyright, citation required, and the terms say nothing on redistribution, so the written confirmation D153 waited on will not be sought. Latinobarómetro forbids republishing and stays out on the same rule; Arab Barometer (registration) and Asian Barometer (application per dataset) were already of no use to set C and are now out on access too.
• Enterprise Surveys microdata (#66). The per-country firm base for customs clearance sits behind login.enterprisesurveys.org. D144's first reopening route, reading the real base from the microdata, is closed. customs_clearance_time stays a check beside Coordination. D144's second route stays open because it needs only published values: a second round on the 2024 questionnaire, read from the World Bank API or a portal table that downloads without an account.
• GESIS/EVS respondent-level pooling (#38, TRUST-1 pooling). The pooled weights need the registered GESIS microdata download. The D64 hold stays on all ten countries with separate EVS and WVS rows: Germany, the United Kingdom, the Netherlands, Armenia, Czechia, Romania, Russia, Serbia, Slovakia and Ukraine. D64's "pooled-microdata treatment" overturn route is closed; a pooling rule that works from the published release tables alone would still qualify.
• Gallup World Poll, which D10 already excluded. volunteering_rate, whose definition Gallup's item matches word for word, stays a declared gap.

(b) Model knowledge is the interpretation layer. An in-session panel is a panel for publication when it has at least three panelists drawn from at least two model vendors, each panelist runs in a separate context with one fixed stance from packages/core/src/delphi/panel.ts, and every panelist scores from the same evidence brief printed by pnpm bench prompt for the current dataset. Its provenance is in_session and is never relabelled gateway: gateway still means real API calls through the AI Gateway (D14). isPanel already accepts such a run, because it counts three or more panel entries on an evidential provenance; what the schema cannot check, context separation and vendor mix, is recorded in the run's note and panel entries (docs/PANEL.md). The invariant does not move: a panel estimate never enters score. It reaches the output only as delphiScore and delphiIqr, and through blendedScore's existing fallback when a dimension has no observed indicator, recorded in blendedFrom. D139 still applies: the run must carry the current datasetVersion before any surface compares it with the indicators. A multi-vendor gateway run remains the stronger instrument, with four vendors, two rounds dispatched by code and failed calls counted, and should be preferred whenever an AI_GATEWAY_API_KEY exists. The two vendors available in session today are Anthropic (Claude, through separate subagents) and OpenAI (through the codex CLI).

(c) The foresight register's spot-check is done by a third model family. D148 asked a person to check a seeded, regime-stratified 20 percent of bodies before the 53 are coded. That check is now done by a model family not used to code: OpenAI through the codex CLI, against the same checklist in docs/research/anticipation/FORESIGHT-REGISTER-SPOTCHECK.md (confirmed, overturned or cannot verify, per body, from the cited sources or their archived copies). D148's thresholds and overturn clause are unchanged: two or more of the 11 sampled bodies overturned, or any body decision that changes a country item, stops the coding.

Why. A source that needs an account or a permission is one the next maintainer cannot reproduce without the same account or the same letter, and one whose derived values the project cannot be sure it may publish. Waiting on those has parked three repairs for weeks with no date. Model knowledge is available now, in session, from more than one vendor; the panel layer was built to hold it beside the indicators rather than in them, so admitting it changes no score.

Cost. Stated with the numbers the closed routes would have moved.

• Trust stays under O1 at 0.398. Afrobarometer would have taken it to 0.410 on set C (docs/research/trust/SURVEY-COVERAGE-FOR-EXPANSION.md), with Latinobarómetro 0.415. Seventeen set C countries keep no A165 value: Angola, Burkina Faso, Botswana, Côte d'Ivoire, Ghana, Guinea, the Gambia, Lesotho, Mali, Mozambique, Mauritius, Malawi, Namibia, Sudan, Tanzania, Uganda and Zambia.
• The guardrail stays at about 0.526 (0.529 on the memo's set C calculation). Afrobarometer would have moved it to about 0.509 and taken survey presence off the income gradient (0.37 to 0.03). Of the 0.526, about 0.07 is the survey gap (0.460 with no survey items anywhere), and that part now has no repair in view.
• Ten countries keep no survey value in Trust, Agency or Shared purpose under the D64 hold. Pooling would have raised Trust to 0.414 but the guardrail to 0.599, so this cost is coverage, not wealth separation.
• Coordination stays at 0.371 against O1's 0.40. The customs row, kept to the 37 countries passing on the proxy base, would have cleared O1 by 0.006 at 7.8.0 (0.406, docs/research/coordination/O1-CANDIDATES.md); all 46 at 2018 or later would have given 0.435. Neither is now reachable through the microdata.
• An in-session panel has fewer independent errors than a gateway run. With two vendors and four stances, each vendor takes two stances, so half the pairs in the IQR share a model's training data; D13's reason for one model per vendor holds and is paid here. Rounds are dispatched by an agent, not code, so a missed panelist is a reviewer's catch, not a counted failed call.
• The foresight spot-check loses its human. Three model families now touch the register (the coders' and the checker's), and none is a person; an error all three share passes.

Overturned by. A source that fails the access test publishing open terms, which readmits it under this rule with no further step (Afrobarometer stating redistribution terms that cover derived country shares is the one worth watching). An owner ruling that accepts one gated source by name. For (b): an in-session panel whose stance-by-vendor spread shows that the two stances on one vendor agree more with each other than with the other vendor's on the same stance, by more than the panel's own median IQR, which would mean the panel is measuring vendors rather than countries and needs a third vendor before publication. For (c): a later human or gateway check that overturns a body the codex check confirmed.

D155: An in-session panel may come from one vendor

Recorded 2026-10-03. Owner ruling of 2026-10-03. Amends D154(b). Extends D13 and D14.

Choice. An in-session panel is a panel for publication when it has at least three panelists, each in a separate context with a distinct fixed stance from packages/core/src/delphi/panel.ts, all scoring from the same evidence brief for the current dataset, even when every panelist comes from one model vendor. D154(b)'s requirement of at least two vendors is dropped. Everything else in D154(b) holds: provenance stays in_session and is never relabelled gateway, the run's note and panel entries name the models, the route and the context separation, a panel estimate never enters score, and D139 still gates any comparison with the indicators on the current datasetVersion. One model may take two stances only in two separate contexts. A smaller model is not added to widen the panel (D13).

A multi-vendor in-session run, or a gateway run, supersedes a one-vendor run as soon as one can be made: it replaces it as the active run on the same dataset version without a further decision.

Why. The only second vendor available in session, OpenAI through the codex CLI, ran out of workspace credits during the first D154 run, after 18 of 125 countries. The two Claude models were available and had already answered round 1. Waiting on a second vendor would leave dataset 9.0.0 without any panel reading beside its 455 thin-coverage cells, which is the gap D154 set out to close.

Cost. The panel's disagreement now measures stance, not vendor. Models from one family share training data, and so share blind spots: a country every Claude model misreads is misread by the median and hidden by a narrow IQR, which then reads as agreement. D13's reason for one model per vendor is paid in full here. The IQR of a one-vendor panel is a floor on real uncertainty, never an estimate of it, and D154's overturn test, which needs two vendors on one stance, cannot be run on it.

Overturned by. A multi-vendor run on the same dataset that disagrees with the one-vendor run's median by more than the one-vendor run's own median IQR on more than a quarter of cells, which would show the shared blind spot is large enough that a one-vendor panel should not publish. Or any gateway or multi-vendor run becoming available, which supersedes the run (not the rule).

D156: The Spanish copy is reviewed by Claude, the chosen terms stay, and four wordings change

Recorded 2026-10-03. Answers #86. Extends D134; changes no address.

Decision. The owner asked Claude to do the review #86 asked a native speaker for, and there is no human reviewer. Claude read every string in packages/core/src/i18n/es.ts and the Spanish words in apps/web/src/lib/words.ts that the [layer] route renders, as an economist from the region would: neutral Latin American Spanish, accents and opening marks, no anglicism where a standard term exists, one word per thing, no advice. 66 strings change. The four terms D134 chose stay:

• Iniciativa (Agency). "Agencia" in Sen's sense is known to development economists but reads as an organisation to everyone else, and no term serves both readers better.
• Ejecución (Building). "Capacidad de ejecución" is the region's standard phrase for a state that turns plans into working systems. It meets "ejecución presupuestaria" in public finance, so the Coordination row that read "Fidelidad en la ejecución presupuestaria" is renamed "Cumplimiento del presupuesto aprobado", which is also not a calque.
• Propósito compartido (Shared purpose), standard and plain.
• Solidez de la evidencia / solidez (confidence) and puntuación (score). "Puntaje" is commoner in Chile, Argentina and Colombia, but "puntuación" is understood in all four and is the standard form.

The map slugs do not move, so no redirect is added and the pinned test in es.test.ts is unchanged.

What does change, beyond wording:

• A documented delivery is a "caso documentado". "Entrega documentada" read as something handed over, and the two other candidates are taken: "iniciativa" is Agency and "ejecución" is Building. Several records are losses or abolitions, which "caso" holds and "logro" would not.
• A map row is read by its position on the scale. The map compares each row's 0 to 100 position with the peer median, and the sentence "Por encima de la mediana de los pares: tasa de homicidios intencionales" read as more homicides when it meant fewer. The three row sentences now open with "Posición en la escala", and the score intro says once that the scale points the way of the capability, so a high position on a row where a higher value means less capability is a low value. Conditions are compared in published units, so their sentences now say "registra un valor mayor que la mediana", which is literally true of informality or transmission losses.
• The agenda's reference lines say what they list. Under a low score, "Puntuaciones utilizables más altas" sat beside the country's number like a model to follow. It now reads "Puntuaciones más altas en esta dimensión, entre las que tienen evidencia utilizable", and "Entregas documentadas en otros países" reads "Casos documentados en otros países". The lines and their place in the document are splitAgenda's and do not change (D39).
• "Benchmark" leaves reader copy, replaced by "la comparación", the word the agenda intro already used. "Sandbox" stays only in parentheses after "espacio controlado de prueba", the term regulators in the region use.

Smaller changes: "¿Qué tan...?" becomes "¿En qué medida...?", which reads neutral in the Southern Cone; "sobre {n} indicadores" becomes "con {n} indicadores"; "contra" becomes "frente a"; the score bands read "media alta" and "media baja"; research grants are "proyectos con financiamiento público concursable" rather than "subvenciones"; "por millón de habitantes" joins "por millón de personas". es.test.ts now pins "caso documentado", the absence of "benchmark", and the row sentences.

Why. #86 asked for a native review before the layers are shown to institutions in the region, and the owner chose a model review instead of waiting for a person. Recording it as Claude's keeps that visible: the D134 overturn clause asks for a Spanish reader, and this is not one.

Cost. No native speaker has read the copy. A model's Spanish is fluent and generic: it will miss what sounds foreign to a reader in Bogotá or Buenos Aires, and it cannot tell which institutional terms an audience in one country uses. Evidence record titles and summaries stay in English inside the Spanish agenda, because they are data, not lexicon. The ground layer's English map has the same backwards row sentence ("Above the peer median: intentional homicide rate") and is not changed here.

Overturned by. A native Spanish reader, ideally in development economics, finding a term or a sentence wrong, which replaces this review with theirs; or a reader of the map taking a "posición en la escala" sentence for the raw value, which would need the row's direction printed beside it.

D157: A World Bank zero is dropped where the row cannot be zero

Recorded 2026-10-03. Resolves A17 as filed; dataset 9.1.0. Extends D25.

Decision. A registry row may carry zeroIsMissing: true. The World Bank ingest drops a published value of exactly 0 on such a row before writing data/observations/worldbank.json, and lists each drop in that run's entry in data/observations/revisions.json under dropped, with the country, year, value, series and the reason zero_is_missing. The diff entry for a cell that left the file this way carries the same reason, so the log never reads the rule as the publisher withdrawing a value. The flag is set on a row from what it measures: a construct a functioning economy of a million people cannot have none of, so an exact 0 can only mean the publisher has no number. It never names a country, and only an exact 0 is dropped; a value of 0.004 is a measurement. The rule applies after the carry-forward of a failed series, so an older file's value passes it too. Denominators, the income context series and checks keep whatever the publisher prints. The pure step is applyIngestRules in packages/core/src/pipeline/ingest.ts, tested in ingest.test.ts.

The audit read every World Bank row that is scored or a condition, 34 rows, against every year and all 125 countries in the file of 2026-10-02. Thirteen rows hold an exact 0 somewhere.

Flagged, with zeros in the file (61 values dropped):

• manufacturing_value_added (% of GDP): Venezuela 1991 to 2011 (21), Brazil 1960 to 1978 (19), the Democratic Republic of the Congo 1991 (1). A manufacturing sector does not vanish for 21 years after reading 14%.
• sci_articles_per_million (a count, divided by population): Serbia 1996 to 2006 (11), the years the articles were counted under Serbia and Montenegro. A country with universities publishes some articles.
• electricity_transmission_losses (% of output): Botswana 1990 to 1997 (8). A grid that carries power loses some of it.
• education_expenditure_gdp (% of GDP): Türkiye 1998 (1). A state with public schools spends on them.

Flagged, with no zero today, because the construct rules one out: rd_expenditure_gdp, researchers_per_million, new_business_density, business_start_days, business_start_procedures, account_ownership, domestic_credit_private, contract_enforcement_days, human_capital_index (a product of terms that are each above zero), labor_force_participation, unemployment_rate, labour_productivity, tax_revenue_gdp and income_inequality. Eighteen rows carry the flag.

Not flagged, with zeros kept, because 0 is a value the construct allows:

• internet_users (% of population): 118 zeros, nearly all 1990, before the internet reached most of these countries; Burundi, Nepal and Sudan into the mid 1990s.
• time_to_export (Doing Business border compliance hours): 78 zeros, 13 EU members 2014 to 2019. Inside the customs union there is no border compliance, and Doing Business records that as 0.
• vocational_secondary_share (% of secondary): 145 zeros, among them the United States and Canada for decades, which had no separately classified vocational track.
• bribery_incidence (% of firms): 7 zeros (Estonia, Sweden, Japan, Lithuania). No surveyed firm reporting a bribe request is a result.
• broadband_subscriptions (per 100 people): 9 zeros, early years in the Democratic Republic of the Congo, Haiti and Lebanon, below the publisher's precision.
• secure_internet_servers (per million): Congo 2010 and Cuba 2011 to 2012. A country can host none.
• tertiary_enrollment (% gross): Botswana 1970 to 1971, before its university admitted its own students.
• high_tech_exports_share (% of manufactured exports): The Gambia 2014 to 2016. A small exporter can ship none.
• electricity_connection_speed (Doing Business distance to frontier, 0 to 100): 56 zeros. The worst performer in the sample scores 0 by construction.

The other rows with no zero today stay unflagged because 0 is possible: statistical_performance and economic_complexity are indices whose floor is a valid reading, budget_execution_fidelity reads 0 for a budget executed as approved, firm_training_incidence is a share of firms, and the three resident filing counts (patents, trademarks, industrial designs) can be 0 where no resident files.

Why. A17. The publisher prints Venezuela's manufacturing value added as 0 for 1991 to 2011, the zero was scored as Venezuela's latest value, and as the row's lowest it set the 0 endpoint for every country (D47). The pipeline imputes nothing, and dropping the value is not imputing: the publisher has no number, and a placeholder read as one is the worse error. A flag on the row keeps the rule a statement about a construct a reader can check, as D120's gate is, and stops it becoming a list of countries whose numbers the project disliked. Logging every drop on every run keeps the rule visible after the first diff (D25).

Dropping the zeros lets the scorer take Venezuela's latest real value, 14.2% in 1990, under the same latest-value rule every row uses; it reads at the recency floor, so Venezuela's Building confidence is unchanged at 0.266. Venezuela's Building goes from 5.9 (125th of 125) to 19.1 (118th). The row's 0 endpoint moves to Lebanon's 1.4% (2021), which normalises to 0 instead of 4.2. 123 current normalised manufacturing values move, every country's but Ireland's, which sets 100, and 123 Building scores move, all but Venezuela's down, by at most 1.4 points (Cuba). Building's correlation with income goes from 0.573 to 0.572, and Building's trend restates for 120 countries because its history is scored on the same frame. No other row moved, and the re-ingest restated nothing else.

The release is dataset 9.1.0, minor under D37: the registry gains a field and the revision log gains dropped and reason, no country is added and no published field is removed, so the frame is not rebased in D47's sense. Scores restate, as they do on any re-ingest, and the changelog says so.

Cost. The judgment is per row and made once: a row added later needs the same audit, and a reviewer may disagree with a flag on a row with no zero today, where the flag changes nothing until one appears. A real 0 on a flagged row, if a publisher ever prints one, is dropped. Venezuela's row now reads a 36-year-old value: the scorer takes the latest observation whatever its age, which A17 now records. The Brazil, Congo, Serbia, Botswana and Türkiye zeros were historical, so they leave the yearly series and the trend and touch no current score.

Overturned by. A publisher documenting a true zero on a flagged row, which removes the flag from that row; a published 0 on an unflagged row shown to be a placeholder, which adds it; or a source-side fix (the World Bank printing Venezuela's series as missing), which leaves the rule in place with nothing to drop.

D158: Brazil's layer opens on what the claim test supports, speaks Portuguese around its pages, and asks for no money

Recorded 2026-10-03. Answers #89. Extends D69, D71, D130, D133, D137 and D138. The Portuguese copy review in it is Claude's, not a native speaker's.

Decision. Brazil's layer is about to be shown to a Brazilian economics institute and a development bank. It changes in seven ways.

1. The opening states what the test supports, computed. /brasil no longer opens on "Riqueza e capacidade são propriedades diferentes", which is the strong claim /thesis reads as failing for what the nine capabilities share. readVerdictPtBr in apps/web/src/lib/verdict-pt.ts reads diagnostics.factorStructure and diagnostics.residualStructure and picks every sentence by the readings D137 and D138 already compute: the factor's income band, the weaker claim's verdict, and the same narrow-margin threshold residual.ts uses (now exported, not copied). No rule is restated and no number is typed. The opening links to a Portuguese verdict section on the same page (/brasil#teste, built by layerVerdictHref), which states the claim in its two strengths and the computed share, income correlation, both verdicts and the income share of a profile, then links to /thesis and /diagnostics in English. On dataset 9.1.0 it reads: one shared factor carries 50.4% of the variation over 123 countries (chance 15.9%, 95th 17.5%), correlates 0.815 with log GDP per capita over 122, so the strong claim does not hold for the shared part; the weaker claim holds, narrowly (shape share 22.7% against 19.0%). 2. The layer describes and never prescribes (D130). The agenda's groups are named for what they hold: "Dimensões com as pontuações mais baixas", "Dimensões com evidência fraca", "Dimensões com as pontuações mais altas", and the front page labels them with RAISE_BELOW, not "Elevar", "Medir antes de gerir" and "Manter". The top-scorer line says it lists the top of the scale, and the evidence-elsewhere line says the records are successes and setbacks outside every score, so neither sits under a low score as a model to copy. splitAgenda and the order are unchanged (D39). A test in pt-br.test.ts rejects "elevar" and "deve" in the lexicon. 3. /brasil/apoie asks for no money. The project is self-funded (owner, 2026-10-03). The page drops the funding section and every funding window it named (Finep, CNPq, the state research foundations, BNDES, IDB, the World Bank, CAF and private foundations), and the link to the ground layer's support page, which lists funded pieces. It asks for four things: use, an outside review of the method (anchor #revisao, built by layerReviewHref), open data and a venue for a seminar. No institution is named in an ask; the data sources a Brazilian series comes from (IBGE, Ipea, Tesouro, CGU) stay. The tab is labelled Participe; the address keeps its slug, so nothing redirects. "A revisão independente está aberta" on the front page now links to that review ask, and "procura parceiros institucionais" with its link to envisioning.com is gone. 4. The page's language is the document's. The root layout's <html> is LangHtml, a client component that reads the path on the server render: a layer's slug gives the layer's language, anything else English. The Spanish layers get es from the same rule. The skip link, the wordmark's name, the header sections, crumbs and tabs, the footer, the search palette and the objection dialog read chromeWords(pathLang(path)) from apps/web/src/lib/chrome.ts, with names of countries, capabilities and indicators from the layer's lexicon. Portuguese has chrome words; Spanish has none yet and keeps the English chrome whole, so no header mixes two languages. This is not a language switch (D69): nothing a page says changes, only the words around it, and only under a layer's slug. The contact form stays English: a Portuguese form would be a second inbox, or a ?lang= on a ground-layer page, and D71 and D69 forbid both. The layer says once that a reader may write in Portuguese. 5. The lexicon reads as Brazilian Portuguese. Claude reviewed every string in pt-br.ts, the Portuguese words in words.ts, the institution summaries and the hand-written copy under apps/web/src/app/brasil, as a Brazilian economist would read them. A score is a pontuação, never a "nota", which reads as a school grade and makes the instrument a report card; the confidence number is solidez (da evidência), because Confiança is Trust; both are pinned by a test. Calques go ("Quão capaz é...", "framework", "o que o país tem para trabalhar"), "ranking" becomes "classificação", a row is placed "na escala". Every indicator now has a Portuguese definition and every unit a Portuguese rendering, which is what left the map's condition descriptions in English; a test holds the lexicon complete over the registry. Global bodies take their Brazilian names (ONU, PNUD, Banco Mundial, FMI, OCDE, OMC, OMS, OIT), state summaries take the right preposition ("no Acre", "na Bahia", "em São Paulo"), and São Paulo's four state bodies no longer fall back to English. Two agenda strings wrote "do {country}" and "em {country}", wrong for most of the 125 countries the Portuguese documents cover; both now name no country, and 62 countries gain the article prose needs. 6. Brazil's evidence records read in Portuguese. Lexicon.evidence maps a record id to a title and a claim, falling back to the record's English through evidenceTitle and evidenceClaim. evidence-pt-br.ts holds all 29 records filed against Brazil. The record stays the sourced English ground layer (D20, D31); a test fails when a Brazil record has no entry, or when an entry's figures differ from the record's in any digit, written the Brazilian way. The records about other countries listed under a dimension stay English. 7. No map sentence reads backwards. A condition is compared in published units, so "o Brasil tem mais do que a mediana dos pares em perdas na transmissão" said more of a bad thing as if it were more of a good one. readCapabilityMap now also returns conditionsLowerBetter, the conditions just compared where a lower value is better, and a lexicon with the optional conditionsLowerBetter string says so after the comparison ("Em perdas na transmissão de eletricidade, um valor menor é melhor."). Portuguese and English carry it; Spanish, whose condition sentences D156 already rewrote as "registra un valor mayor", does not yet. Row sentences now say a row's position on the scale is above the peer median, in Portuguese and in English, which is the change D156 made in Spanish and listed as owed in English.

No university is named in the layer's copy: the new-export-products caveat named Harvard in Portuguese and English and now names the Growth Lab's trade data. Spanish still names it and is the Spanish review's to change.

Why. The layer goes before readers who will check the opening against the thesis, read a recommendation into "elevar", and read a funding window as a request for money from them. Each fix makes a page say only what the data and the decisions say. Recording the review as Claude's keeps the D130 and D134 overturn clauses honest: no Brazilian has read this copy yet.

Cost. No native speaker has reviewed the Portuguese. A model's Portuguese is fluent and generic; it will miss what reads as foreign to a reader in Brasília or Rio, and the term choices (pontuação, solidez, Agência, Adaptação) are judgment. The chrome now depends on the path, so a second chrome language is a table, and a page served outside a layer's slug in its language would get English chrome. The contact form, the method, limits, decisions, objections and subnational pages, and every evidence record not about Brazil, stay English inside a Portuguese reading. The evidence translations are hand-written copies of sourced claims: the digit test catches a changed figure, not a changed meaning. Dropping the funding section removes the only place the layer said what a sponsor would buy; the ground layer's /support still lists funded pieces and is not changed here.

Overturned by. A Brazilian reader, ideally an economist, finding a term or a sentence wrong, which replaces this review with theirs; a reader of the opening taking it for the strong claim, which would cut it back to the verdict section's link alone; the owner reopening outside funding, which restores an ask without naming windows; or a Portuguese contact form accepted by a decision that supersedes D71's one inbox.

D159: A value more than 15 years old does not count toward a score

Recorded 2026-10-03. Owner's decision; dataset 9.2.0. Amends D22 and D45 in how a cell is observed; extends D47 in what sets the frame.

Decision. A cell whose latest value is more than MAX_SCORED_VALUE_AGE (15) years before the reference year is treated as missing for the score, the frame and coverage. The reference year is the currentYear the scoring run is given, which the CLI sets to the calendar year the dataset is generated: in a 2026 release the oldest year counted is 2011, and 2010 is set aside. The constant and the test, isTooOldToScore, live in packages/core/src/pipeline/confidence.ts beside the confidence bands; the rule names no country and reads only the year.

How it meets the rest of the model:

• Score and coverage floor (D45). A set-aside cell is not observed. It leaves the mean and observedIndicators, so a dimension that falls below MIN_INDICATORS_FOR_SCORE publishes no score. The Delphi fallback (D63) still needs zero observed rows, so a dimension left on one row reads none.
• Confidence. Coverage falls, because the row stays in the denominator as an unobserved one. Recency and source quality average only observed cells, so a set-aside value leaves them; a value at 13 to 15 years already sat on the 0.1 recency floor, so the age limit is where that floor stops counting at all.
• Frame (D47). A set-aside value is not on the ruler: it sets no Tukey fence and no endpoint. Every country is still scored against the frame, which is now every country's latest value inside the limit.
• Trend (D22, D24). Momentum already reads each end with a five-year maximum age, so the limit does not change what a trend reads; it moves only through the frame. Each row's yearly series stays published in full, set-aside years included, normalised against the current frame, so an old value may clamp there.
• The row says so. The indicator row reads missing with staleExcluded: { year, raw }, and the viewer prints the year and why it does not count on the country page and on the capability map, through Lexicon.capabilityMap.rowStale in every lexicon. The per-indicator files under data/out/indicators/ list only scored cells, so a set-aside value is absent there as it is from the frame.
• Conditions and checks are kept and marked. Neither is scored, so the limit has nothing to remove them from. A condition is what a country has to work with and a check sits beside the score; an old value is still the publisher's value, and the reader decides what a 1990 enrolment rate says. Each carries stale: true and the viewer prints its age. A condition keeps its rank among the countries with a value.

Measured against dataset 9.1.0, with no re-ingest (the observation file is unchanged):

• 25 scored cells are set aside: Shared purpose 11, Experimentation 6, Building 3, Agency 2, Adaptability 1, Trust 1, Learning 1. 45 condition values and three check values are marked.
• Nine scores are withdrawn in eight countries, each now on one row: Experimentation for Burkina Faso, the Republic of the Congo and Mali; Shared purpose for Guinea, The Gambia, Papua New Guinea, Trinidad and Tobago and Venezuela; Learning for Venezuela. Scored countries: Experimentation 125 to 122, Shared purpose 123 to 118, Learning 124 to 123.
• Mean confidence: Experimentation 0.318 to 0.317, Shared purpose 0.319 to 0.317; the other seven unchanged.
• Correlation with log GDP per capita: Agency 0.466 to 0.445, Experimentation 0.728 to 0.719 (n 123 to 120), Building 0.572 to 0.571, Shared purpose 0.259 to 0.250 (n 122 to 118); the other five unchanged.
• The guardrail (mean confidence against log GDP, D117) moves from 0.526 to 0.527 (n 123).
• D137: the first factor's share goes from 0.504 to 0.494, on 115 complete countries instead of 123, and its correlation with income from 0.815 to 0.814.
• D138: every verdict holds. The residual structure share is 0.309 (0.308), on 115 complete cases (122); peers still differ, the release check still reads mixed, leave-one-out still robust, the income share still little, and the weak claim holds.
• 178 dimension scores move, 393 normalised cells with them. The largest: Azerbaijan's Shared purpose 46.7 to 23.9 (its 2005 Gini set aside); Malawi's Agency 37.7 to 56.6 and Haiti's 14.0 to 21.1 (new business density from 2009 and 2010 set aside, leaving the two 2019 Doing Business rows, A14); Indonesia's Shared purpose 66.2 to 84.3 and Iran's 55.0 to 72.8 (tax revenue from 2009); Jordan's Shared purpose 44.8 to 31.3; Venezuela's Trust 35.0 to 23.8; Cuba's Adaptability 64.6 to 55.8 (its 2010 long-term unemployment share, so Cuba joins A16).
• The three cases that prompted it. Venezuela's manufacturing share, 14.2% in 1990, is set aside, so its Building goes from 19.1 (118th) to 8.9 (124th) on two rows (A17). Mali's one patent filing in 1981 is set aside, so its Experimentation, 0.1, is withdrawn. The Gambia's 1990 tax revenue is set aside, so its Shared purpose, 54.0, is withdrawn.

The Doing Business rows frozen at 2019 (A6, A14) are inside the limit until the 2034 release, so the rule does not touch them now.

The release is dataset 9.2.0, minor under D37: fields are added (staleExcluded on an indicator row, stale on a condition and a check), no country is added and no field is removed, so the frame is not rebased in D47's sense even though which values set it changes. Scores, coverage and confidence restate, and the changelog says which.

The panel run in data/delphi/latest.json is for dataset 9.0.1, and isDelphiRunForDataset requires an exact match, so since 9.1.0 no panel estimate is published beside a score. That is unchanged here and is not this decision's doing; it needs either a run on the current dataset or a decision on how far a run may carry across a minor release.

Why. A value scored at any age said the same thing as a current one in the score, and only the recency part of confidence knew the difference. A 1981 patent count, a 1990 tax ratio and a 2005 Gini describe a different country. Fifteen years keeps every Doing Business row (2019) and every survey round of the last decade and a half, and drops what no reader would quote as current. A cut on the year alone is a rule a reader can recompute from the published files and stops the choice becoming a list of countries.

Cost. Coverage falls where publishers stopped, which is mostly in poor countries: of the eight countries that lose a score, six are low or lower-middle income. The guardrail barely moves, because each loses one cell, but the factor test and the residual structure now run on 115 complete countries instead of 122 and 123. A country with one old value on a row that a newer source would fill reads empty until that source exists, where before it read something. The limit is a hard edge: a value from 2011 counts fully in a 2026 release and a value from 2010 does not, and every year the line moves forward one year, so a re-score without a re-ingest restates when the calendar year turns.

Overturned by. Evidence that a set-aside value describes the country as well as a current one on some row (a slow-moving construct with a stable series), which would set a per-row limit in the registry rather than raise this one; or a reader of a withdrawn score showing the old value told them something the empty dimension does not, which would publish the old value on a separate, unscored line.

D160: A panel estimate carries across a release of the same major version only where the evidence it read is unchanged

Recorded 2026-10-03. Owner's decision; dataset 9.3.0. Supersedes the exact-match clause of D139 (isDelphiRunForDataset requiring the run's datasetVersion to equal the current one); D153 and the rest of D139 stand.

Decision. A panel run applies to a dataset in two ways.

• Same dataset version: every estimate applies, as before.
• Same major version and same prompt version: each estimate applies only if the evidence brief the current dataset builds for its country and dimension hashes to the briefHash stored with the estimate. A cell whose brief moved is dropped, and so are both of its rounds. A run with no stored hashes carries nothing across versions: nothing says what its panelists read, so it stays a research note until it is rerun.
• A different major version never carries. A major version rebases the frame (D47), the same sentence can sit on a different ruler, and D139 and D153 hold that no estimate crosses one. A different prompt version never carries either.

What is hashed. dimensionBrief in packages/core/src/delphi/prompts.ts is the block of the evidence brief that carries one dimension: its heading, its question, the coverage line (coverage, mean age weight and the indicator-derived score) and one line per indicator, each with the value as printed, the year, the normalised value and the source. cellBriefHash is the SHA-256 of that text. The country header, the stance, the round, the round-1 summary and the scoring instructions are left out: they say nothing about the evidence, and the cell key already names the country. The same text is the substring of both the round 1 and the round 2 prompt, which a test holds. Indicator judgements are matched the same way: indicatorAuditHash covers a dimension's name, question and audit rows, stored as auditHash, so a judgement carries only while the rows it judged are unchanged.

Because the brief prints the normalised value and the indicator-derived score, a hash changes when the frame moves under an unchanged value. That is intended: the panelist read a number on a ruler, and the ruler moved.

How it is applied. applyDelphiRun (packages/core/src/delphi/apply.ts) is the only place the decision is made. scoreAll scores once without the run to build the briefs, applies it, and reads delphiScore, delphiIqr and the blendedScore fallback from what applies. bench score writes the applied run, with an application record (cells held, cells carried, each dropped cell and why, judgements kept and dropped) to data/out/delphi-applied.json, and the viewer reads that file and never data/delphi/latest.json, which stays the run as produced. The record reaches diagnostics.delphiApplication and the report. The hash and the record are optional fields, so runs written before this decision still validate. runDelphi stores the hashes on every new run. The 9.0.1 run's hashes were backfilled by rebuilding its briefs from the 9.0.1 published files at commit 15dbad08; the 375 round-1 prompts staged for that run came out byte for byte the same as the ones rebuilt there.

Result on 9.3.0: 348 of the 455 cells carry, and 107 are dropped. Eighteen dropped because the cell's own row was set aside by D159 (Shared purpose 11, Experimentation 5, Building 1, Learning 1). Eighty-eight because a row the panelist read keeps its value and sits at a different normalised value on the moved frame: 84 Experimentation cells on resident patents, trademarks and industrial designs per million (the D159 set-asides in Burkina Faso, the Republic of the Congo, Mali, Ethiopia and Guatemala move those scales), and four Building cells on manufacturing value added (D157 drops Venezuela's placeholder zeros and D159 sets aside its 1990 value and Sudan's 2009 one). One, China on Experimentation, changed only in its recency weight.

Why. The exact-match rule was right for a frame rebase and wrong for a release that restates a handful of rows: since 9.1.0 the whole panel was withdrawn, 455 cells, because 107 of them had read evidence that changed. The panel is an interpretation beside the indicators, and what it is shown is the brief. A cell that read the same brief is the same judgement, and a hash of the brief is a test that can be recomputed from the published files.

Cost. A carried estimate was made on a release that is not the current one, and the viewer's run line says which. A hash is exact, so a one-point move of a normalised value drops a cell whose judgement a panelist would not have changed; the rule prefers dropping to guessing. A frame move that changes few values still empties a dimension of panel estimates, as Experimentation shows. Backfilled hashes rest on the published 9.0.1 files matching what the panelists saw, which the byte comparison confirms for round 1 and not for round 2, whose briefs come from the same code.

Overturned by. A carried estimate that a rerun of the same panel on the current brief moves by more than the panel's own IQR, in cells whose hash matched, which would mean the brief hash leaves out something the panelists read; or a run on the current dataset, which supersedes the carry.