What changed

Release notes for the benchmark frame and the viewer that publishes it. App releases describe the whole user-facing product; dataset releases describe the scored data contract. The build checks that the newest entry in each stream matches its source version.

The entries before App 1.0.0 are retrospective, high-level summaries reconstructed from the repository history. They document explicit version bumps and may skip versions that were never committed.

App 1.27.0: 2026-10-03

• Panel estimates are back beside the scores. 348 of the 455 panel cells scored on dataset 9.0.1 apply to dataset 9.3.0, because the evidence each was made on is unchanged. The 107 whose evidence moved are dropped and listed in the report and the diagnostics. The panel page says how many cells were carried. Decision D160 sets out the rule.

Dataset 9.3.0: 2026-10-03

• A panel estimate now carries across a release of the same major version when its evidence is unchanged. Each panel estimate stores a hash of the evidence brief it was made on (briefHash, and auditHash on an indicator judgement), and an estimate applies only if the current brief hashes the same. Nothing carries across a major version. Fields are added and none is removed, so no score moves. diagnostics.delphiApplication records what carried and what was dropped, and data/out/delphi-applied.json holds the estimates that apply. Decision D160.
• 348 of 455 cells carry from the 9.0.1 run. The 107 dropped are 90 Experimentation, 11 Shared purpose, five Building and one Learning, from the rows D157 and D159 changed and the frame moves they caused.

App 1.26.1: 2026-10-03

• A row set aside as too old says so. On a country page and on each capability map, a row whose latest value is more than 15 years old shows that value, its year and that it does not count. Conditions and behavioral checks that old show their age.
• Known limits A1, A14, A16 and A17 are restated for the age limit, and the glossary defines it.

Dataset 9.2.0: 2026-10-03

• A value more than 15 years old no longer counts. In this release the oldest year that counts is 2011. An older latest value is left out of the score, the scale and coverage, and the row shows it with its year. It stays in the row's history. Conditions and behavioral checks keep old values and mark them. Decision D159 sets out the rule and its cost.
• 25 rows are set aside in 21 countries, 11 of them on Shared purpose and six on Experimentation.
• Nine scores are withdrawn in eight countries, which now have one row each: Experimentation for Burkina Faso, the Republic of the Congo and Mali; Shared purpose for Guinea, The Gambia,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.
• 178 scores move. The largest are Azerbaijan's Shared purpose (46.7 to 23.9), Malawi's Agency (37.7 to 56.6), Indonesia's and Iran's Shared purpose (66.2 to 84.3 and 55.0 to 72.8) and Venezuela's Building (19.1 to 8.9), which no longer reads a 1990 manufacturing share.
• What income explains barely changes. Agency's correlation with income goes from 0.466 to 0.445 and Experimentation's from 0.728 to 0.719. The first factor's share goes from 0.504 to 0.494, now on 115 countries with all nine scores, and every reading of what is left after income holds.
• Indicator rows gain staleExcluded, and conditions and checks gain stale. Same 125 countries and the same observations as 9.1.0.
• Model panel estimates are still not shown, because the panel run is for dataset 9.0.1.

App 1.26.0: 2026-10-03

• Brazil's layer opens on what the test supports. The Portuguese front page no longer says wealth and capability are different properties. It says what the current release shows: what the nine capabilities share follows income, and countries at similar income have different capability shapes, by a narrow margin. Every number is computed from the release, and the full verdict is now on the page in Portuguese.
• The layer describes and does not advise. The agenda groups capabilities by what they hold, lowest scores, thin evidence, highest scores, instead of "raise", "measure first" and "keep". The participation page asks for use, an outside review of the method, open data and a seminar venue, and no longer lists funding programmes: the project is self-funded. Its invitation to review the method now leads somewhere.
• The Portuguese pages speak Portuguese around them too. The page language, the header, the footer, search and the objection form are in Portuguese on Brazil's layer. Brazil's 29 documented deliveries, every indicator description and the names of international bodies are now translated, and a score is a "pontuação" throughout. The Spanish layers now declare Spanish as the page language. Claude reviewed all the Portuguese copy; no native speaker has yet. Decision D158 sets out the choices.
• Map sentences no longer read backwards. Where a lower value is better, as with power lost in transmission, the map now says so after comparing a country with its peers, in Portuguese and in English.

App 1.25.9: 2026-10-03

• Known limit A17 now describes what is left. Venezuela's manufacturing row no longer reads the publisher's zero. It reads 1990, the last year before the zeros, and the limit says how old that is.
• The glossary defines a placeholder zero.

Dataset 9.1.0: 2026-10-03

• A published 0 is dropped where the row cannot be 0. Eighteen rows, among them manufacturing value added, public education spending and electricity transmission losses, measure something every working economy has some of. On those rows a World Bank value of exactly 0 is a missing number, so the ingest now drops it and lists each drop with its reason in the revision log. Rows where 0 is a real result keep their zeros: border time inside the EU, internet use in 1990, the share of firms asked for a bribe. Decision D157 lists every row and why.
• 61 values were dropped, all of them on four rows: Venezuela's manufacturing share for 1991 to 2011, Brazil's for 1960 to 1978 and the Democratic Republic of the Congo's for 1991; Serbia's scientific articles for 1996 to 2006; Botswana's transmission losses for 1990 to 1997; and Türkiye's education spending for 1998.
• Building scores restate. Venezuela's zero was the bottom of the manufacturing scale for every country. Lebanon's 1.4% now sets it, so 122 other countries' manufacturing values and Building scores move down, Building by 1.4 points or less. Venezuela's Building goes from 5.9 to 19.1 (125th to 118th of 125), because its manufacturing row now reads 14.2% from 1990. Building's trend restates with them, and nothing else moved.
• Revision log runs gain a dropped list, and a revision a rule caused carries a reason. The registry gains a zeroIsMissing flag. Same 125 countries.
• Model panel estimates are no longer shown. The panel run was made on dataset 9.0.1, and a run is shown only beside the dataset it was made on.

App 1.25.8: 2026-10-03

• A panel from one model vendor can now be published. Three or more panelists, each with its own stance and working on its own, may all come from one vendor. Their spread then reads the stance and not the vendor, and a run with a second vendor replaces it when one can be made. Decision D155 sets out the cost.
• A new known limit, A17. The World Bank prints Venezuela's manufacturing value added as 0% of GDP for every year from 1991 to 2011. The zero is scored and sets the bottom of that row's scale for every country.

Dataset 9.0.1: 2026-10-03

• Model panel estimates now sit beside 455 scores. Three Claude panelists, Opus 5.5 twice and Sonnet 5.5 once, each holding one stance, scored every capability where published data cover half the rows or fewer, in all 125 countries, over two rounds. Their median and spread are published next to the score and never change it.
• The panel reads Trust lower than the indicators. On the 52 countries with a Trust score it sits 20.5 points below on average while ranking them in nearly the same order (rank correlation 0.83). On Shared purpose the order differs more (0.51). On Anticipation and Experimentation the panel stays within eight points and keeps the order (0.95 and 0.93).
• No cell shows dissent. The widest spread between panelists is 8 points. With one vendor that spread is a lower bound on the uncertainty.
• Scores, confidence and ranks are unchanged from 9.0.0. Cuba's Agency and Shared purpose, which have no observed row, now show the panel's estimate as their fallback reading.
• The panel prompt now prints per-head rows per head. It used to show a national total under a per-million unit, and a rejected row as an empty value.

App 1.25.7: 2026-10-03

• The Spanish pages for Mexico, Colombia, Chile and Argentina read more plainly. 66 strings were revised for neutral Latin American Spanish. The capability names and the words for score and confidence stay. On the map, a row now reads by its position on the scale, so a low homicide rate no longer reads as sitting above the median. Documented deliveries are now "casos documentados". Claude did the review, not a native speaker; decision D156 records it.

App 1.25.6: 2026-10-03

• The benchmark now uses open data only. A source joins only if anyone can download it without an account and republish what is derived from it. Three repairs that needed a registration or a permission are closed: Afrobarometer's trust question for 17 African countries, the firm counts behind the customs clearance check, and the pooled survey rows for the ten countries whose survey values are held. Trust keeps a mean confidence of 0.398, under the 0.40 target, and the link between confidence and income stays at 0.526. Decision D154 sets out the cost.
• A panel of models from two vendors can now be published. Three or more model panelists, each with a fixed stance and working on its own, scoring from the same evidence, count as a panel. Their estimates sit beside the scores and never change them. No panel has been run on the current dataset yet.

App 1.25.5: 2026-10-02

• 15 more documented deliveries for the new countries. Four document a loss: Belgium's strategic mask stock, Jordan's work permits for Syrian refugees, Austria's Bildungskarenz and Bulgaria's Belene nuclear plant. The flagships include Rogun, BelAES, Denmark's hospital building programme, Czechia's D1 modernisation and Belgium's BE-Alert. A second reviewer checked every number at its official source. The capability agenda is regenerated.

App 1.25.4: 2026-10-02

• 220 more documented deliveries for the countries that joined the frame, 24 of them losses, and 48 searched notes. 746 of the 1,000 country and gap cells now carry a record or a note. Among the losses are ended rail and water concessions in Tanzania, Zambia, Guinea and Mozambique, and Denmark's technology board. A second reviewer opened every source and checked every number, corrected 107 records and dropped 18. The capability agenda is regenerated.

App 1.25.3: 2026-10-02

• The first documented deliveries for 16 more of the new countries. 28 records reach Mozambique, Romania, Ghana, Kazakhstan, Madagascar, Tunisia, Côte d'Ivoire, Sri Lanka, Nepal, Zimbabwe, Mali, Zambia, Burkina Faso, Guinea, Malawi and Burundi. Seven document a loss, among them the Beira railways concession, Nepal's Melamchi water supply, Zambia's solar milling plants and Burundi's coffee sector partnership. The flagships include Mozambique's disaster agency, Romania's A1 motorway, Sri Lanka's Southern Expressway and Ghana's Sankofa gas project. A second reviewer checked every number at its official source. The capability agenda is regenerated.

App 1.25.2: 2026-10-02

• The first documented deliveries for 14 of the new countries. 28 records reach Pakistan, Italy, Bangladesh, Ukraine, Morocco, Russia, Egypt, Uzbekistan, the DR Congo, Iraq, Iran, Angola, Tanzania and Uganda. Seven document a loss, among them Pakistan's Neelum-Jhelum hydropower plant, Italy's fintech sandbox, Iran's targeted subsidies and adult literacy programmes in Bangladesh, Tanzania and Uganda. The flagships include the Padma Bridge, MOSE, Noor Ouarzazate, Julius Nyerere hydropower and Ukraine's eRecovery. Nothing for Myanmar or Sudan passed the inclusion rule. A second reviewer checked every number at its official source. The capability agenda is regenerated.

App 1.25.1: 2026-10-02

• The limits page now reads the 125-country frame. Every known failure on it is checked against dataset 9.0.0, with ranks out of 125 and the country lists that apply now.
• Trust's lower link to income is explained. It fell from 0.606 to 0.354 because the poorer countries added lack the two survey rows, not because Trust came apart from income. On the same rows everywhere it would be 0.579.
• Experimentation now names the countries that file patents and trademarks at a regional office, where the national counts are thin or decades old.
• The Adaptability limit covers 15 countries without a long-term unemployment share, and the Coordination limit names ten states that coordinate through the executive. Both show on those countries' map pages.
• Across the 125, closed regimes do not read calm on polarization as a group, so that limit now names the single autocracies among the calmest countries.

Dataset 9.0.0: 2026-10-02

• The benchmark now covers 125 countries. It adds 72 to the 53 it had: every country of a million people or more that has enough published data to score all nine capabilities (D153). Among them are Italy, Belgium, Austria, the Nordic countries not yet in it, most of Central and Eastern Europe, Russia, Ukraine, Pakistan, Bangladesh, Egypt, Iran, Iraq and 22 countries in Africa south of the Sahara. All 72 publish nine scores.
• Every score restates, and 8.3.0 numbers are not comparable with these. Each indicator's scale is set by every country in the benchmark, so adding countries moves it for everyone. Of the 470 scores the 53 had in both releases, 462 moved, by 3.2 points on average. The largest rises are in Experimentation for Israel (28.3 to 48.0), Sweden and the United Kingdom, because the countries added below them widen the bottom of the patent and trademark scales. Ranks are now out of 125.
• Brazil: Anticipation 44.6 (51st of 124), Agency 58.0 (54th), Coordination 89.1 (21st), Trust 53.4 (73rd), Learning 36.6 (73rd), Experimentation 28.7 (38th of 125), Adaptability 64.5 (62nd), Building 38.3 (43rd) and Shared purpose 30.6 (118th of 123).
• Trust drops below the 0.40 confidence target, at 0.398. Half its rows come from one survey release that reaches few of the poorer countries added, so their Trust scores rest on court and contract rows alone. Experimentation falls from 0.364 to 0.318 and Shared purpose from 0.343 to 0.319. Nothing was lost for the 53: their confidence is unchanged.
• How much the capabilities follow income has shifted. Trust's correlation with income falls from 0.61 to 0.35 and Agency's from 0.58 to 0.47. Experimentation rises from 0.65 to 0.73, Building from 0.43 to 0.57 and Adaptability from 0.43 to 0.53. The one shared factor behind the nine carries half their variation, and it correlates 0.815 with income, with a 95% interval of 0.745 to 0.867. Countries at the same income still have different capability shapes, by the rule fixed before the first test, and income now accounts for less of the typical country's profile than it did on 53.
• Confidence now tracks income more: 0.526, against 0.272. The original 53 happened to include the poor countries one large survey reached; the world's data is thinner where incomes are lower. The 0.526 is the new baseline the project watches, and a source that raises it inside these 125 countries is the warning sign.
• The new countries arrive without evidence records, foresight coding or panel estimates. Their 576 evidence cells are open and listed as the research queue. Belarus joins China and Cuba among the countries whose GitHub figure is held, because its count of public projects barely grew while the platform grew elsewhere.

Dataset 8.3.0: 2026-10-02

• National belonging is back on the list of things Shared purpose asks for (D151). It was dropped at 8.0.0 because the only survey item aimed at it, national pride, reads highest in electoral autocracies. That item is still not used, and pride is no longer part of the definition, but the question of whether people count themselves part of the national community stays open as a gap, the same way trust in institutions and polarization already did.
• Shared purpose confidence falls from 0.41 to 0.34, below the 0.40 target again. Nothing was lost: the dimension rests on the same three rows it had before 8.0.0, and the higher figure came from dropping a question while no new data arrived. In 28 countries Shared purpose moves from usable to thin evidence, so their agendas now list it under measure instead of raise. China's leading item becomes Anticipation, and Switzerland and South Korea have no raise item left. No score moves.
• Four gaps now ask for something people or institutions do (D152). Venture capital becomes first venture deals per million people, counted by deal instead of by amount. Adult digital skills becomes the share of adults who did a named digital task in the last three months. Regulatory sandbox activity becomes firms admitted to and completing a sandbox per million people. Long-horizon research becomes the share of new public research grants that run five years or longer. All four are still gaps, so no confidence or score moves. Their ids are unchanged.
• The rule behind both changes: a question the benchmark still wants answered stays a gap when the only data aimed at it fails, and a row is retired only when the question itself is dropped or was never about a capability. The glossary entries for gap and retired indicator say so.
• Same 53 countries and the same scores as 8.2.0. The guardrail, mean confidence against income, moves from 0.274 to 0.272.

Dataset 8.2.0: 2026-10-02

• Adaptability gains a row: new export products (D149). It counts the goods a country barely exported in 2009-2011 and exported competitively in 2022-2024, out of the goods it had room to enter, from Harvard's Growth Lab trade data at a fixed release. Entering new lines is an economy moving people and money into things it was not making. Poland leads at 8.3%; Brazil entered 20 of 872, 2.3%, 34th of 53.
• The nine countries that rested on two rows now have three. South Korea, India, Mexico, Peru, Uruguay, China, the Philippines, El Salvador and Haiti still have no long-term unemployment figure, but their Adaptability confidence rises from 0.38 to 0.46. Mexico falls from 88.0 to 64.3 (6th to 28th) and El Salvador from 86.5 to 62.6 (8th to 32nd); India rises from 10th to 4th and China stays 7th. A16 is restated: the unemployment rate in those nine still has nothing beside it in the score that tells a fluid market from one where nobody can afford to stay unemployed.
• Every Adaptability score moves, because the new row is a quarter of each. The United States falls from 88.4 to 71.6 and Brazil from 73.7 to 61.4 (35th to 37th); the United Arab Emirates rises from 68.7 to 72.0. Mean confidence rises from 0.52 to 0.58 and the correlation with log GDP per capita falls from 0.46 to 0.43. No other capability moves.
• Informal employment is published beside Adaptability, not scored (D150). The ILO's share of work with no contract or social protection reads 87% in India, 71% in Peru, 64% in El Salvador and 57% in Mexico, which is why a low unemployment rate there is not a fluid market. 43 of 53 countries have a value.
• The shared factor across the nine scores carries 0.50 of their variance, unchanged to two places. Same 53 countries; 8.1.0 Adaptability scores are not comparable with these.

App 1.25.0: 2026-10-02

• 72 new country pages, each with its nine scores, its agenda and its capability map, and all of them in the country list, the comparison picker, the front-page field and the sitemap.
• The Spanish and Brazilian layers name every country in their own language where they mention one.
• Indicator notes that quoted a coverage count from the old 53 now say so, or give the count on 125.

App 1.24.11: 2026-10-02

• The evidence grid is complete. Every country now has a record or a searched note for every delivery gap. Ecuador's adult learning cell closes on a note: the national school register counts whole institutions, so a fall in adult enrolment cannot be told apart from adults moving into ordinary schools. Cuba's housing programme is now a documented loss, from 20,232 homes in 2022 to 5,493 in 2025, filed like Venezuela's and Singapore's housing programmes. Two Brazilian records move to university and industry collaboration, and two records filed against the retired business research row leave the corpus (D147).

App 1.24.10: 2026-10-02

• A new known limit, A16: Adaptability rests on two rows in nine countries. South Korea, India, Mexico, Peru, Uruguay, China, the Philippines, El Salvador and Haiti have no long-term unemployment figure, so their Adaptability score is the unemployment rate and export concentration alone. Mexico, China, El Salvador and India sit in the top ten on that basis. A low unemployment rate can mean people cannot afford to stay out of work, and on two rows nothing checks it. The capability map names A16 on those nine countries' Adaptability pages.
• A13 now quotes Shared Purpose's current correlation with income, 0.202.
• Country pages no longer log a radar mismatch in the browser console. The chart's coordinates are rounded the same way on the server and in the browser.

App 1.24.9: 2026-10-02

• 64 more documented deliveries, 15 of them losses. Every country now has a record or a searched note for every delivery gap but one, adult learning in Ecuador. The losses include the UK's private finance initiative, the US child tax credit expansion, Germany's dual apprenticeships, Australia's Inland Rail and Brazil's adult education enrolment. Policy Horizons Canada and Sweden's commission on the future replace two notes that had closed those countries' foresight cells by mistake. Five older records whose sources had moved now point at the new address or an archived copy. A second reviewer opened every source and checked every number. The capability agenda is regenerated.

Dataset 8.1.0: 2026-10-02

• Experimentation gains a row: new public software repositories per million people (D145). It counts how many public projects people in a country started on GitHub in the year to the first quarter of 2026, from the GitHub Innovation Graph at a fixed commit. A new repository costs nothing to start and is kept when it is abandoned, which is the many cheap attempts the dimension asks about. Brazil adds 18,446 per million, 17th of 51.
• China and Cuba have no value on the row. Their repository counts barely grew (down 0.1% and up 2.8%) while the median country's grew 24.6%, which reads access to the platform. A published rule holds any country under a quarter of the median growth, and a held country is never scored at zero.
• Experimentation's mean confidence rises from 0.31 to 0.36, still under 0.40, and its correlation with log GDP per capita rises from 0.57 to 0.65. Ireland rises from 4.8 to 19.7 and the Netherlands from 28.1 to 40.2; Germany falls from 81.7 to 68.2 and Turkey from 64.0 to 50.7. Brazil stays at 25.6. No other capability moves.
• The one shared factor across the nine scores carries 0.50 of their variance, up from 0.49. Same 53 countries; 8.0.0 scores in other capabilities are unchanged.

Dataset 8.0.0: 2026-10-02

• Scores restate in Learning and Adaptability, and 7.x numbers are not comparable. Three rows that record what a country has, not what it does, leave the scores and are published beside them as conditions: the vocational share of secondary school (Learning), and labor force participation and electricity transmission losses (Adaptability) (D141).
• Adaptability now reads unemployment, long-term unemployment and export breadth. Its correlation with income falls from 0.74 to 0.46 and its mean confidence from 0.64 to 0.52. Countries with low unemployment and broad exports rise: India from 59.5 to 85.4, Mexico from 69.2 to 88.0, Honduras from 43.8 to 70.5. Switzerland falls from 75.3 to 65.3. Nine countries, India and Mexico among them, read it on two rows. Brazil reads 73.7.
• Learning now reads the Human Capital Index, firm training and citation impact. Singapore leads at 84.2, the United Arab Emirates rises from 49.9 to 73.6 and Bolivia falls from 57.2 to 35.9. Brazil reads 35.0. Cuba no longer has enough rows for a Learning score.
• Two declared gaps are retired. Business share of R&D (D142) reads how a country's research spending is split, and counts state enterprises as business. National belonging (D143) could only be read through national pride, which runs highest in electoral autocracies. Neither moves a score. Both raise their dimension's confidence without a new observation: Experimentation from 0.27 to 0.31, Shared purpose from 0.34 to 0.41.
• Customs clearance time is shown beside Coordination and not scored (D144). Firms report how many days their exports took at customs, in 50 countries. The number of firms behind each figure is not published, and the survey changed its question in 2024.
• The one shared factor across the nine scores now carries 0.49 of their variance, down from 0.52. Same 53 countries.

App 1.24.8: 2026-10-02

• Known limits brought to dataset 7.8.0. The Trust entries now cover the two survey items side by side. China, Japan, South Korea, Singapore and Malaysia say most people can be trusted but few trust a stranger, and their Trust scores fall on the new row. Court compliance is now the row that carries most of Trust's tie to income.

Dataset 7.8.0: 2026-10-02

• Trust now reads trust in strangers. A new Trust row measures the share of adults who trust people they meet for the first time completely or somewhat, from the same Joint EVS/WVS release as generalised trust, in 37 countries. It ranges from 7.8 in Ecuador to 73.9 in Sweden, and Brazil reads 22.7. Countries where people trust "most people" but not a stranger fall: Japan from 79.4 to 64.3 on Trust, Singapore from 84.6 to 70.8, South Korea from 80.8 to 67.6 and China from 67.6 to 55.8. Ethiopia rises from 41.9 to 45.6. Trust's mean confidence goes from 0.35 to 0.42 and its correlation with income from 0.675 to 0.606. The value adds two published columns, which is the one place the benchmark sums a survey's categories (D140). Germany, the United Kingdom and the Netherlands are held, as for generalised trust. Same 53 countries.

App 1.24.7: 2026-10-02

• The polarization limit is restated on V-Dem's 2026 release. Closed autocracies still read calmer than electoral regimes, though no longer level with the liberal democracies. The United Arab Emirates and Cuba now carry the pattern, while Rwanda and Singapore no longer read calm. Scored, the item would still lift the Emirates by 11 points on Shared purpose, so it stays a check.

App 1.24.6: 2026-10-02

• 26 more documented deliveries, eight of them losses. Foresight units and regulatory sandboxes are now covered in every country: each either has a record or a note saying what was checked and why nothing qualified. The losses include the US Office of Technology Assessment, Finland's SHOK research clusters, Poland's government strategy centre, the US consumer finance sandbox and the Dutch STAP training budget. The flagships include France's CIFRE industrial doctorates, Switzerland's Innosuisse projects, Germany's parliamentary technology assessment office and Mexico's adult literacy institute. A second reviewer checked every number at its official source and dropped two records. The capability agenda is regenerated.

App 1.24.5: 2026-10-02

• Known limits restated on the current dataset. Coordination no longer reads small states low: Uruguay scores 73.9 and Costa Rica 88.6, and the score is unrelated to population. What remains is narrower, and A9 now says so: the civil society row pulls down states that coordinate through the executive, Singapore most of all. Every other entry is refreshed to 7.7.1 or names the run its figures come from. No limit quotes the old one-panelist estimates any more, because they were scored on a different frame (D139). A7 is now about Korea and Japan reading low on Learning.

Dataset 7.7.1: 2026-10-02

• V-Dem moves to release 16 (March 2026). Court compliance and civil society strength now read 2025 values, and the polarization and turnout checks follow. V-Dem re-estimated its whole series for this release, so 2024 values restate as well as gaining a year: 52 of 53 court compliance values and 51 of 53 civil society values moved. Trust's correlation with income goes from 0.671 to 0.675 and Coordination's from 0.563 to 0.558; mean confidence is unchanged in both. The largest country moves are South Africa down 3.4 on Trust and Haiti up 6.4 on Coordination. Same registry, same 53 countries.

App 1.24.4: 2026-10-01

• 31 more documented deliveries, nine of them losses. Every country now has at least three. The losses include HS2, Mexico's disaster fund FONDEN, Singapore's Tuaspring desalination partnership, PPP Canada, China's PPP programme, India's UDAN regional flights and Sweden's health emergency stockpile. The flagships include Crossrail, the Champlain Bridge, the Hokuriku Shinkansen, Singapore's Jobs Support Scheme, Ireland's research centres and Nicaragua's disaster insurance through CCRIF. A second reviewer checked every number at its official source. The capability agenda is regenerated.

App 1.24.3: 2026-10-01

• 27 more documented deliveries, six of them losses. The losses are Peru's southern gas pipeline, Bolivia's Mutún steel plant, Thailand's disaster warning towers, Germany's Bildungsprämie, Spain's fintech sandbox and Spanish employer training. The flagships include Peru's Talara refinery, Indonesia's 35,000 MW programme, Ecuador's Coca Codo Sinclair, Germany's LNG terminals, Argentina's Atucha II, Poland's Vistula Spit canal and Rwanda's Marburg response. A second reviewer checked every number at its official source. The capability agenda is regenerated.

App 1.24.2: 2026-10-01

• 30 more documented deliveries, eight of them losses. The new records reach 16 more countries. Among the losses are the US Affordable Connectivity Program, Cuba's currency unification, Australia's VET FEE-HELP loans, France's pandemic mask stock and Montevideo's Rivera rail line. Among the flagships are the Interstate Highway System, Operation Warp Speed, the Swiss shelter system, Finland's Report on the Future, Estonia's state stockpiles and Chile's infrastructure concessions. A second reviewer checked every number at its official source. The capability agenda is regenerated.

App 1.24.1: 2026-10-01

• 25 documented deliveries for the 16 countries that had one. Each of them now has two to four. Ten of the new records document a loss, among them Kenya's adult education centres, Canada's long-gun registry, China's zero-COVID policy and Haiti's PetroCaribe projects. The flagships include China's high-speed rail, the Barakah nuclear plant, Tel Aviv's Red Line and Cuba's hurricane civil defence. Every number was read at an official source and checked a second time by a separate reviewer. The capability agenda is regenerated.

App 1.24.0: 2026-10-01

• The thesis tests what is left after income. "Where the claim holds, and where it fails" now reads in three steps: what the nine capabilities share (the one-factor test of D137, which looks like income), what is left once income is taken out, and what that means for the strong claim (capability separate from wealth) and the weaker one (countries at the same income have different capability shapes). Every verdict is a template the computed reading chooses, with the n beside it. On dataset 7.7.0 the strong claim does not hold for the shared part and the weaker one holds on 50 countries, on a narrow margin the page states. The per-capability wealth chart stays.
• /diagnostics gains a section with all four tests, their nulls, the rule that reads each and the release and leave-one-out table. The front page's income module gains one sentence, and the glossary defines "capability shape". No country's residual is published anywhere. See D138.

Dataset 7.7.0: 2026-10-01

• New published field: diagnostics.residualStructure (D138). Four aggregate tests on the wealth residual over the countries with all nine residuals: whether the residuals share a factor (against D137's chance level and a permutation baseline), whether income peers differ in shape (against incomes dealt at random and residuals dealt at random), whether each dimension's residual order holds between consecutive releases (read from git) and when one country is dropped, and how much of a country's profile income accounts for. Each carries its reading under rules fixed before the first run, and the weaker claim's verdict. Only statistics over countries: residual.json is unchanged and remains the only file with per-country residuals. schema/residual-structure.schema.json describes the field. No score or confidence changes.

App 1.23.1: 2026-10-01

• Israel's desalination joins the documented deliveries. Private concessions supplied a third of the country's fresh water in 2022, according to a State Comptroller audit that also records the weak oversight behind it. The capability agenda is regenerated.

App 1.23.0: 2026-10-01

• The thesis shows whether the nine capabilities are one thing. "Where the claim holds, and where it fails" gains a computed reading of the one-factor test: how much of the variation one shared factor carries (52.9% over 51 countries), what chance would give at that size (18.9%, 21.5% at the 95th percentile), and how closely the factor follows income (r 0.86, n 50). At that strength the page says plainly that the shared factor looks like income. A chart draws the share against the chance band and the factor's correlation with income at every dataset release since 1.0.0. /diagnostics publishes the loadings, the eigenvalues, the countries left out and the full release table, the front page's income module gains one sentence, and the glossary defines "first factor". See D137.

Dataset 7.6.0: 2026-10-01

• New published field: diagnostics.factorStructure (D137). The correlation matrix of the nine dimension scores over the countries with all nine scored, its eigenvalues, the first factor's share and loadings, its correlation with log GDP per capita, and a seeded Monte Carlo chance level. Below 30 complete cases the dimensions scored for at least 90% of countries are solved as well. data/out/factor-history.json holds the same test at every committed dataset release, read from git by bench diagnose, and schema/factor-structure.schema.json and schema/factor-history.schema.json describe both. No score or confidence changes.

App 1.22.1: 2026-10-01

• Two sandbox records. Brazil's central bank admitted seven of 52 applicants to its first sandbox cycle. India's central bank admitted 28 entities across four themed cohorts and found 13 products viable in the three whose exits it published. The capability agenda is regenerated.

App 1.22.0: 2026-10-01

• Every country has a capability map in English. /country/<ISO3>/map lists the nine capabilities with the score and the confidence as two numbers and the median of the 10 countries nearest in income, and /country/<ISO3>/map/<dimension> reads one capability: the indicators the score rests on, the conditions beside it and the country among its peers on the field chart. It is the same computed reading as the Brazilian and Spanish layers' maps, for all 53 countries, in the ground layer's English. It is the Map tab in each country's pages, beside the agenda. On a country with a layer, the header's language switch moves between the two maps on the same capability. The sitemap lists the 530 new pages and /llms.txt points at the pattern. See D136.
• Facts about one country's row come from one table. The few things a map says about one country's indicator that the published data does not carry, such as the survey behind a long-term unemployment series or why the row is empty, are now one table read by every language, held to the pinned ILOSTAT release by a test. The ground-layer map shows them for every country they are true of, 14 in all. The Brazilian and Spanish pages read as before.
• English names that take an article get one. The agenda and the map write "the United States", "the Netherlands", "the United Kingdom", "the United Arab Emirates", "the Philippines" and "the Dominican Republic" inside a sentence. The six English agenda documents are re-rendered. No score changes.

App 1.21.0: 2026-10-01

• The evidence corpus has a finish line (D135). Every country is checked against eight gaps a documented delivery can speak to, 424 cells in all. A cell closes with an evidence record or with a dated note of what was searched and why nothing qualified. 142 cells are closed today. The glossary gains an entry for the grid, and the research queue no longer offers a cell that is already closed.

App 1.20.0: 2026-10-01

• Mexico, Colombia, Chile and Argentina get Spanish layers. /mexico, /colombia, /chile and /argentina each hold an overview, the agenda and the capability map with its nine pages, in Latin American Spanish, read from the same files as the English pages. The overview is computed: the nine scores with the solidez de la evidencia beside each, links to the map and the agenda, and the known artefacts that bear on the country's capabilities. The layers hold no institutions, states or support pages, because the project has not done that work for these countries. Each country's English profile offers its Spanish reading in the header, the sitemap lists every new page, and the feed carries the four Spanish agendas. Brazil's pages are unchanged. See D134.

App 1.19.1: 2026-10-01

• Portuguese copy names confidence "solidez da evidência". In pt-BR, "Confiança" was both the Trust dimension and the label for the 0 to 1 evidence number, so the Trust map page used the word twice in two senses (the cost D133 recorded). The evidence number is now "solidez da evidência", shortened to "solidez" in table headers, chips and the agenda's item headings, across the lexicon, the Portuguese glossary, the Brazil pages and the rendered .pt-BR agendas. Trust keeps "Confiança". English copy is unchanged. Copy change only; no score or confidence changes.

Dataset 7.5.0: 2026-10-01

• Confidence in the courts is published beside Trust as a check (D132). The Joint EVS/WVS share answering "a great deal" for the justice system, 37 countries. It is not scored: closed and electoral autocracies read far higher than democracies on every confidence item tested (Vietnam and China above 85% on "a great deal or quite a lot"), and it does not follow what governments do with court rulings (r -0.13 with court compliance). The institutional_trust gap stays open. No score or confidence changes.

App 1.19.0: 2026-10-01

• Brazil's layer maps all nine capabilities. /brasil/mapa lists the nine in the model's order, each with Brazil's score, its confidence and where the score sits against the median of the 10 income peers, and links to one page per capability at /brasil/mapa/<name>, computed by the same buildCapabilityMap as D130's Adaptability page. A capability with no conditions (Coordination, Trust, Experimentation, Shared purpose) says so and draws no conditions panel; thin confidence is stated beside the score. Each page names the known artefacts that bear on its capability, by id, from one table checked against the limits document. The layer's tab strip gains one Mapa tab in place of the Adaptação tab, and /brasil/adaptacao redirects to /brasil/mapa/adaptacao. The sitemap lists all ten pages. See D133.

Dataset 7.4.0: 2026-10-01

• Trust scores court compliance (D131). V-Dem's coding of how often the government complies with important court rulings it dislikes, 53 of 53. It records what the state does when a court rules against it, which is what "strangers cooperate on the strength of the rules" rests on. Trust now scores 52 countries, mean confidence 0.31 to 0.35, and its correlation with income rises from 0.57 to 0.67, reported as a finding. The row tracks regime type closely (r 0.88 with V-Dem's electoral democracy index), and it pulls down Trust where bribery under-reporting had lifted it: China 88.7 to 67.5, Rwanda 78.9 to 63.9.
• The V-Dem sweep declines every other candidate. Bribery and corruption items read reputation, the construct D23 retired; deliberation and common-good items read closed autocracies as good (A13); administration and fiscal items are conditions. Coordination gains nothing from V-Dem. See docs/research/vdem-sweep/TRIAGE.md.

Dataset 7.3.0: 2026-10-01

• Experimentation scores industrial design applications (D126). Resident design filings per million people, World Bank IP.IDS.RSCT, 50 of 53. A filed design is a registered attempt at a new product form, cheaper and more frequent than a patent. Experimentation now scores all 53 countries, mean confidence 0.23 to 0.27, and its correlation with income falls from 0.62 to 0.57. It partly overlaps trademarks (r 0.82).
• The GEM extension is held (D125). GEM's 2019 to 2025 reports would take entrepreneurial activity and fear of failure from 16 countries to 40, but the 14 they miss are mostly lower-income, and confidence would come to track income (0.29 to 0.39). The values are recorded in the research memo and not scored.

App 1.18.0: 2026-10-01

• Brazil's layer maps Adaptability. /brasil/adaptacao reads Brazil's Adaptability in Portuguese: the score and its confidence, the five rows it rests on with their published values, the broadband condition beside it with its correlation with income and with the score, and Brazil's place among its 10 income peers on the flag field. Every sentence compares a value with a peer median, and the page recommends nothing. It is computed by buildCapabilityMap from the published files and regenerates with every release. The field chart and the conditions panel take Portuguese words, and the glossary defines capability map and income peers. See D130.

Dataset 7.2.0: 2026-10-01

• diagnostics.json gains income. The latest GDP per capita (PPP, constant international dollars) for each country, with its year, from the series the wealth tests already read. It is context and enters no score. It is published so the capability map's income peers are computed from the data (D130).
• Agency scores perceived control (D127). The Joint EVS/WVS mean answer to how much freedom of choice and control people feel over their lives, 1 to 10, for 37 countries. It is a perception and labelled as one. Agency's mean confidence rises from 0.38 to 0.48 and its correlation with income falls from 0.64 to 0.58.
• Shared purpose scores charitable membership (D128). The share mentioning membership of a humanitarian or charitable organisation, the same release and 37 countries. Shared purpose now scores 51 countries instead of 47, mean confidence 0.26 to 0.34, and its correlation with income falls from 0.46 to 0.20.
• Voter turnout is published beside Shared purpose as a check (D129). V-Dem's turnout of registered voters at each country's latest election, 52 of 53. Compulsory voting moves it by about 17 points and regime type barely at all, and it reads the democratic channel A5 retired, so it is not scored.
• Confidence tracks income less. Mean confidence against log GDP per capita falls from 0.33 to 0.29.

Dataset 7.1.0: 2026-10-01

• Trust scores bribery incidence (D123). The share of firms asked for a bribe in public transactions moves from a check to a scored row in Trust's institutional family. It records experience, not reputation, and D60 held it out only for its income correlation, which D118 no longer allows. Trust now scores 50 countries instead of 37, mean confidence 0.21 to 0.31, and its correlation with income falls from 0.61 to 0.57.
• Learning scores research citation impact (D124). The share of a country's articles and reviews in OpenAlex's top 10% for their field, as a ratio to the world, covers 53 of 53. Learning's mean confidence rises from 0.37 to 0.50 and its correlation with income from 0.66 to 0.75.

App 1.17.0: 2026-10-01

• Conditions sit beside every capability. Country pages list what the country has to work with under each capability, with the published value, the year and its rank, in a panel separate from the behavioural checks. Capability pages show how each condition goes with income and with the capability's score. The Brazil layer and both agenda lexicons carry the same reading in Portuguese and English. /sources and /indicators mark conditions, /explore draws them as squares, and the glossary defines the term. See D122.

Dataset 7.0.0: 2026-10-01

• Ten bought conditions leave the scores (D122). Research spending, researchers and secure servers (Anticipation), internet users, account ownership and private credit (Agency), tertiary enrolment and education spending (Learning), broadband (Adaptability) and output per worker (Building) are fetched and published as conditions on each dimension, with a rank and never a 0 to 100 value. They enter no score, confidence or trend. diagnostics.json gains conditions: each one against income and against its dimension's score.
• Scores move in five dimensions and are not comparable with 6.x. Agency falls from r 0.85 to 0.64 against log GDP per capita, Learning from 0.73 to 0.66, Adaptability from 0.84 to 0.74 and Building from 0.64 to 0.43. Anticipation stays at 0.87. The first factor's share of the dimension scores falls from 0.62 to 0.50. Agency and Learning confidence fall below 0.40. Brazil reads 45.8 on Anticipation, 51.6 on Agency, 30.2 on Learning, 65.4 on Adaptability and 28.2 on Building.

App 1.16.1: 2026-10-01

• Three evidence records, one of them a reversal. The Gotthard Base Tunnel opened on schedule at 51 percent over its 1998 cost basis. Finland has shelter places for about 4.8 million people. Germany ran its public shelter places down from about 1.6 million to 477,593, none of them operational. The capability agenda is regenerated on the larger corpus.

App 1.16.0: 2026-10-01

• Checks can come from any source. The sources page gains a database or release column for behavioural checks and prints the pinned file and variable behind a check that does not come from the World Bank. The first is political polarization from V-Dem. See D121.

Dataset 6.2.0: 2026-10-01

• A row is chosen for what it measures (D118). A candidate is decided on its construct. Its correlation with income is published as a finding and no longer passes or fails it. Dimension correlations with income are reported, not targeted.
• Adaptability gains two observed rows. Export concentration from UNCTADstat covers 53 of 53 countries (D119). The long-term unemployment share from ILOSTAT covers 44, behind a plausibility gate that drops questionnaire artefacts by rule, not by country name (D120).
• Political polarization is published as a check, not scored. V-Dem's v2cacamps reads closed autocracies as calm, and low polarization under repression is not shared purpose. It sits beside Shared purpose with that reason (D121, A13).

App 1.15.1: 2026-09-24

• Reader copy says what each page holds. Headings, ledes, the llms.txt summary and the English and Portuguese lexicons drop metaphor, slogans and internal facts, and follow American spelling. No number, score or dimension changes. Agenda markdown keeps its old intro lines until the next agenda run.

Dataset 6.1.2: 2026-09-23

• The business R&D gap says why it is still a gap. Its note claimed UIS publishes the series. UIS stopped in 2023, the remaining sources cover too few countries, and together they track income over the screen. No value changes.

App 1.15.0: 2026-09-23

• The sources page prints every call. Behavioural checks are fetched on every ingest and never scored; /sources now lists them with the request that fetches each one, so the page repeats every call the benchmark makes. See D116.
• Charts and type follow the scale. Every chart line reads a named stroke weight, the last text-sm is gone, and the agenda history axis reads "0 to 100". Internal links all come from one set of helpers.

Dataset 6.1.1: 2026-09-23

A correctness fix to the out-of-frame flag and a regeneration of three stale research layers. No score and no confidence changes.

• A historical value beyond every current value is flagged. When a current outlier was winsorized onto its fence, a historical value further out clamped without setting outOfFrame. Four more cells now count as clamped in the momentum baskets. See D115.
• Residual, velocity and leverage cover 53 countries. The three offline layers had not been regenerated since Portugal joined the frame.
• Docs quote the current dataset. KNOWN-ARTEFACTS, the research roadmap and the panelist brief now state 6.1 figures, recomputed from the output.

Dataset 6.1.0: 2026-08-31

Retired indicator rows leave the coverage denominator. Scores and the normalization frame are unchanged; published confidence values restate upward for Coordination, Trust, Building and Shared Purpose.

• A rejected dataset is no longer counted as a hole. Coverage is now observed indicators over the observed and gap rows. A row retired for cause stays in the registry, stays published with status: "retired" and stays in the diagnostics, but no longer lowers confidence. See D100.
• Two dimensions cross a band. Coordination and Shared Purpose move from very thin to thin. Trust stays very thin: it rests on two observed rows.

App 1.14.0: 2026-09-06

• Coverage is drawn as the count it is. How many of a capability's indicator rows were observed for a country used to read as a two decimal ratio, and as the words "not measured" where the count fell below the floor. It is now one tick per row, filled where the row was observed, with a notch at the two rows a score needs. The count is still printed beside it.
• An observed year sits on a rail. Every indicator record and every publisher row now shows when the value was observed as a position between 1990 and this year, with the line trailing the tick showing how long ago that was.
• A workbench page collects every mark. Not published and in no navigation: it draws each chart beside the states it can take, from the live index.

App 1.13.0: 2026-09-02

• The front page shows two drawings instead of naming them. The wealth module carries the lane field in its measured arrangement, and a new module carries Brazil's system matrix. Each picture is the link to its page. See D113.

App 1.12.0: 2026-09-02

• The dot motif behind the front page breathes. The bubbles grow and shrink on the parent brand's 3.2 second cycle, at the same faint alphas as before. A reader whose system asks for reduced motion gets the still frame, and the motion stops while the band is off screen. See D112.

App 1.11.0: 2026-09-02

• The institution network sidebar is a facet navigator. Systems, levels and jurisdictions now have their own index and detail destinations beside the institution directory. Their cards show the number of institutions they contain, while the institution list keeps exact filters and group-by controls. See D111.

App 1.10.0: 2026-09-02

• The institution network has a readable scope. Broad network views open with state-level institutions hidden, because Brazil's feed has 282 of them against 75 federal institutions. A route-owned switch adds them back without losing the current URL filters, and institution-specific views keep their state relations visible by default. The full multi-select Level filter stays available for more exact combinations. See D110.

App 1.9.1: 2026-09-02

• The institution network chrome is solid and deterministic. The Envisioning panel aliases now resolve inside the route-owned theme, and the dense network toolbar no longer lets graph lines bleed through its controls.

App 1.9.0: 2026-09-02

• The indicator registry can be read as lanes. A new page at /explore draws every indicator in the lane of the capability it measures. A filled dot has data, a dashed ring is a declared gap and a thin ring is a retired row. A second arrangement moves each indicator to how closely its series tracks income, with the wealth threshold drawn as a rule and each capability's own correlation as a tick. A line joins the indicator pairs the diagnostics find redundant. Pointing at a dot reads it under the field, and the registry and the diagnostics pages both link to the drawing. See D108.

App 1.8.0: 2026-09-02

• Bodies no country owns have a place. A global ledger at data/institutions/global.json holds the UN, its programmes and agencies, the multilateral lenders and the intergovernmental standard setters, each once, with the benchmarked countries that are members of it. A country map reaches a body by id in a sourced relation, and the body then appears in that country's institution pages and drawn network at the global level. Brazil's map reaches UNDP, the IMF and the UN today.
• Membership is a fact about the body. Each global profile states how many of the benchmarked countries belong to it and whether the country being read does. Nothing in the ledger enters a score or a confidence. See D107.

App 1.7.0: 2026-09-01

• The institution map can be read as a network. A new page at /network draws Brazil's institutions and the relations between them, opening on one institution's neighbourhood rather than on the whole graph. Both institution pages link to it.
• The relation ledger is still where a relation is read. A drawn line shows that two institutions are connected and which family the relation belongs to. Which way it runs, and the verb that names it, stay on the institution page.

App 1.6.3: 2026-08-31

Panel run path, ahead of the first gateway run.

• A run no longer activates itself. A full-frame run used to replace latest.json whether or not --activate was passed, which put an unreviewed panel behind every Delphi surface the moment it finished. Activation is now always explicit.
• Failed provider calls are counted and published. A call that fails after retries is still dropped so the run survives, but attemptedCalls and failedCalls now travel on the run file, the command prints a warning, and bench validate reports both the failure count and any cell-round that came back with fewer panelists than the panel declares.
• The default panel is four distinct vendors. Mistral joins Anthropic, OpenAI and Google. The previous three-model default dealt one vendor two stances.
• Panel prices come from the gateway. The pricing table is read from the gateway's public model list, which states the rate a run is billed at, so no panel model carries an unverified price. See D106.

App 1.6.2: 2026-09-01

• The charts draw from one set of weights and shades. The radar, the flag field, the flag bubble and the sparkline carried nine stroke widths and twelve opacities between them, each chosen in its own file. They now read one scale of five weights and five shades, so a hairline means the frame and a heavy line means the data on every chart. Nothing moved by more than 0.1.
• A design check runs before every build. It warns, and never fails, when a stroke width is written by hand, when text-sm appears in a viewer file, when an em dash reaches reader copy, or when a repository document is named as a bare path instead of a link.

App 1.6.1: 2026-08-31

• Confidence values are compact. Every confidence value uses one band-colored numeric chip instead of a separate meter and number, while the exact value and shared band legend remain visible.

App 1.6.0: 2026-08-31

• Search is available everywhere. The top navigation opens a searchable command palette with ⌘K/Ctrl+K, covering pages, countries, capabilities and indicators.

Dataset 6.0.0: 2026-08-31

Portugal (PRT) joins the benchmark as its 53rd country. The normalization frame is rebased and every published score is regenerated; Dataset 5.1.0 values are not comparable with this release.

• Portugal is source-backed from the existing pipeline. The release adds World Bank history, a Joint EVS/WVS trust observation and a V-Dem civil-society observation without adding synthetic or manual values.
• Portugal enters the research queue. Five documented, non-scored evidence records now cover four of its gap indicators, while the research inventory keeps 21 uncovered country-gap slots visible for further source work.

App 1.5.6: 2026-08-31

• Challenge entry is centralized. Score tables and radar readouts no longer repeat a challenge button; the top navigation now opens one shared form where readers choose the country and capability they want to contest.

App 1.5.5: 2026-08-31

• Hover menus switch cleanly. Leaving one top-level item closes its menu immediately, so moving across the navigation never leaves two panels open.

App 1.5.4: 2026-08-31

• The top navigation hides while scrolling down. It returns when the reader scrolls up, reaches the top, navigates to a new page or focuses the header.

App 1.5.3: 2026-08-31

• The NCB lockup follows the parent brand alignment. The glyph now sits beside the wordmark, with the full benchmark name aligned beneath it.
• Hover explanations are quieter. Scores, confidence, trends and chart points now use visible legends, inline context and accessible labels instead of a field of native browser tooltips. The intentional country field hover card remains.

App 1.5.2: 2026-08-31

• The contextual navigation joins the header. The subnav now shares the top-level navigation's surface and outer rule, with a tighter rhythm between the active section and the links it contains.

App 1.5.1: 2026-08-31

• Country shapes appear on the homepage field. Hovering a flag in the distribution now shows its nine-capability radar; the Countries menu remains focused on navigation.

App 1.5.0: 2026-08-31

• Agenda evidence is positioned on the timeline. Dated agenda items now appear on an aligned event rail for the selected capability, with the full item list kept visible below it.
• Longer history is the default view. The chart opens on the longest published capability span available for that country.

Dataset 5.1.0: 2026-08-31

The historical source window now reaches 1960, and agenda evidence carries its documented start year into the published agenda JSON.

• The source request stays complete at the longer horizon. World Bank requests allow the full 52-country historical response instead of truncating once the 1960 window exceeds the former page size.
• Long spans remain selective. Ten-, twenty-, thirty-, fifty- and oldest-window momentum entries are attempted; only matched baskets that pass the existing evidence floor are published.

App 1.4.0: 2026-08-31

• Agendas show capability history. Each country agenda now has a selectable dimension chart with ten-year, twenty-year and longer spans where the matched evidence supports them.
• Long history stays honest. The chart keeps the current 0–100 frame, reports the matched-indicator count and leaves unsupported spans absent rather than interpolating them.

Dataset 5.0.1: 2026-08-31

The historical observation window now starts in 1976, adding the older source values needed to test approximately 50-year capability movement.

• Longer spans are published when earned. The score output now attempts ten-, twenty-, thirty- and oldest-available-year momentum entries.
• Current scores remain the same contract. Older observations feed history and do not enter the current frame or headline dimension score.

App 1.3.1: 2026-08-31

• Country navigation uses one contextual band. The country context, reading choices and page tabs no longer occupy separate horizontal rows.

App 1.3.0: 2026-08-31

• The agenda shows its research blind spots. A complete 52-country by nine-capability matrix keeps every empty cell visible and links each filled cell to its source-checked delivery records.
• Counts remain inventory, not scores. Row and column totals describe the evidence corpus only; they never enter capability scores or confidence.

Dataset 5.0.0: 2026-08-31

The Brazil subnational pilot is now a maintained, self-describing diagnostic layer.

• Reconciliation is computed. The former aggregate claim for the state Gini is now independent, with the equal-unit recomposition and signed residual published beside the national coefficient.
• The output contract is generic. Files live under subnational/{ISO3}/{indicatorId}.json, are indexed, and are generated from a registry rather than a one-off adapter.
• Subnational values remain outside the benchmark. They never enter a national frame, score, confidence, agenda or ranking. Aggregate claims are rejected when their per-series tolerance is exceeded.

App 1.2.2: 2026-08-31

• Brazil's local reading names the diagnostic clearly. The page now shows the computed recomposition and residual without describing the independent state Gini as corroborating a national value.

Dataset 4.5.0: 2026-08-31

Coordination gained a full-frame, adapter-backed civil-society measure.

• V-Dem civil society strength is now scored. The pinned Country-Year Core v15 release contributes v2x_cspart for all 52 benchmark countries at 2024.
• The adapter and provenance are reproducible. The source release, variable, year and archive URL are fixed in the catalog; the derived observation file carries the expert-coded source tier and license note.
• The frame stays at 52 countries. This is a minor release: the new row changes scores and confidence, but does not rebase the country ruler.

App 1.2.1: 2026-08-31

• The footer lists what the header opens. Its columns are now the site's sections, read from the same navigation tree the header walks, so every capability appears there alongside every method page. The footer had drifted: Capabilities opened nine pages in the header and offered one at the foot of the page.

App 1.2.0: 2026-08-31

Sections open their pages from the header, and the header stays with the reader.

• A section can be opened without being entered. Hovering a section in the header shows the pages under it, so what Countries, Capabilities, Method, Participate and About hold is reachable from anywhere instead of only from inside. The section is still one control: a click goes there. On a phone, where there is no hover, the menu sheet opens a section in place.
• The Countries menu shows the country you are reading. Open it from a country page and it draws that country's shape above the links, the same nine-axis radar the countries grid uses. Away from a country the menu is unchanged.
• The menus close the way a reader expects. Moving the pointer away, Escape with focus back on the section, or arriving at a new page. For a keyboard, ArrowDown opens a section and the arrow keys walk it, with Home and End reaching its ends.
• The front page opens on what the benchmark is testing. The band now states the claim it can fail, that a country's capability is separate from its wealth, instead of describing the shape of the data. The dimension count, the common scale and the confidence beside each score are all still there, in the sections under it that draw them.
• The dot motif on the front page is quieter. The band's texture was competing with the sentence it sits behind, so the dots now read as atmosphere rather than as something to look at.
• The header and the tab strip travel together. Both are pinned to the top of the window, so a section is one click away at any depth of a long page.

App 1.1.0: 2026-08-30

• Brazil's institution map is published for a drawn network. The map is projected into a feed at /api/institutions/BRA, in English and in Portuguese, so the network can be drawn on a surface of its own. The relation ledger on the institution page is still where a relation's direction and its verb are read.
• Only a jurisdiction the map has actually mapped is drawable. The union, the state of Sao Paulo and the municipality of Sao Paulo carry enough recorded relations to have a shape. The 26 state entries that are still scaffolds do not, and the feed says so rather than drawing them.

App 1.0.1: 2026-08-30

• The viewer has an icon. The Envisioning mark now sits on a rounded near-black tile in the browser tab, on a home screen and in a bookmark list, generated at every size the platforms ask for.
• The front page opens on a dark band. It spans the window, on the same surface the footer uses, patterned with the dot motif the rest of Envisioning draws behind a hero.
• The navigation hangs from one edge. The sections, the trail into a page and the tabs under the header now line up on the right on a wide screen.
• Buttons, fields and filters share one shape. Every control on the site now comes from one place, so a form, a filter strip and a dialog agree about how tall a button is and how large its label reads.

App 1.0.0: 2026-08-30

The first formal product release gathers the benchmark, its research surfaces and its public viewer into one navigable application.

• The viewer is a complete reading surface. Country profiles, comparisons, capability pages, agendas, diagnostics, sources and method pages are linked through one navigation tree.
• Research is visible without being confused with scores. Brazil's institutional map and subnational reading, the thesis, evidence workflow and provisional leverage, velocity and residual layers each state their scope.
• The project is ready to be used and challenged. Contact/support paths, embeddable views, social metadata, an agent-readable contract, public feeds and a human changelog are part of the release.
• This app release publishes Dataset 4.4.0. The dataset remains the stable reference for any quoted score; this app version identifies the surrounding product release.

Dataset 4.4.0: 2026-08-29

The Trust dimension gained its first source-backed international adapter while the 52-country frame stayed fixed.

• Trust now has a source-backed adapter. The pinned Joint EVS/WVS A165 table enters the observation store as publisher-weighted data.
• Behavioural evidence remains separate. The bribery-incidence check stays visible beside Trust but remains outside scores, confidence and coverage.
• The existing 52-country frame remains the reference. No country was added and no normalization rebase was performed in this release.

Dataset 4.3.0: 2026-08-29

This release made the boundary between scored evidence and useful but non-scoring checks explicit.

• Behavioural checks became a first-class output. Bribery incidence is published beside Trust with its reason for exclusion, but does not enter a score, coverage floor or confidence calculation.
• Trust was split by evidence family. Social and institutional evidence are kept distinct, with court-case clearance recorded as a gap for future work.
• The release added fields without moving published scores. The version jump records a contract change while leaving the existing country results comparable.

Dataset 4.1.0: 2026-08-28

Coordination gained a directly observed measure of whether public budgets are executed close to plan.

• Budget execution fidelity was added as a scored indicator. The pipeline uses a distance-from-100 transform so both under-spending and over-spending count as deviation.
• The frame was rescored and the new indicator was published in the country, indicator and flat-table outputs.

Dataset 4.0.0: 2026-08-28

The benchmark expanded to a complete 52-country frame and rebased its ruler.

• Twelve Latin American countries completed the registry, bringing the published set to 52 countries and completing the region's coverage.
• Every score was restated against the new frame. Because each country helps define the normalization endpoints, this was a major, announced dataset change; old scores are not directly comparable.
• The Portuguese country reading covered the full region, alongside the viewer and method-page updates that explained the rebased frame.

Dataset 3.0.0: 2026-08-28

The scoring frame stopped treating any country as an external reference.

• All countries were scored against a ruler they helped build. Tukey fences and the 0–100 endpoints now come from the full country set in the frame.
• Every published score moved as a result, so the dataset received a major version bump.
• The panel workflow became inspectable. A second institutionalist panelist, chat-ready prompts and merged panel replies were added to the research workflow.

Dataset 2.0.0: 2026-08-27

The benchmark began withholding claims where the evidence could not support a dimension score.

• Wealth-correlated evidence was retired where diagnostics showed it was misleading, rather than allowing it to stand in for capability.
• A coverage floor was introduced. Thin dimensions publish their indicator rows, confidence, trends and evidence gaps, but no numeric dimension score.
• The viewer learned to preserve missingness honestly, leaving unsupported radar axes empty and labeling them as not measured.

Dataset 1.1.0: 2026-08-27

The first data contract was regenerated with the provenance and trend details needed to inspect a result over time.

• Trend outputs gained clamp counts and matched-basket context.
• Diagnostics gained out-of-frame and wealth-attribution fields.
• Data gaps and agenda holds became explicit published states, rather than disappearing into an empty result.

Dataset 1.0.0: 2026-08-27

The first versioned benchmark contract established the project as a published, inspectable dataset rather than only a prototype viewer.

• The benchmark model covered nine capability dimensions with equal-weight dimension scores and confidence reported separately.
• World Bank observations, evidence gaps and Delphi estimates were kept distinct, with provenance carried through the pipeline.
• The output became self-describing. A semantic version, JSON Schemas and a Frictionless Data Package were emitted alongside country, indicator and table outputs.