Skip to content
NCBNational Capability Benchmark
Navigate with ↑ ↓Enter to openEsc to close

Challenge the benchmark

Choose a score and give the benchmark a specific reason to reconsider it.

When you submit, the form attaches the current score and confidence for the country and capability you chose.

Submissions are public and start as awaiting review. A maintainer can add a response and signature after review.

What this benchmark is

NCB is a research prototype from Envisioning. It scores 125 countries on nine capabilities using public data, gives every score its own confidence value, and publishes no overall ranking.

What is in the current release

How big the dataset is, how much of it is measured, and when it last ran.

Dataset
9.3.0

Adding a country rebases every score and bumps the major number.

Countries
125

All of them set the frame that every score is measured against.

Indicators
34 of 75

18 are declared gaps. They stay in the registry and lower confidence.

Generated
2026-10-03

Every published number comes from this run.

Raw values carry their source and their year. The pipeline drops missing values and never fills them in, so a thin dimension shows low confidence. Why any of this is worth measuring is the thesis, which also shows how much of the claim the data supports.

What it will not do

Four things it does not do.

  • No overall score. An average of the nine capabilities would hide where a country is strong and where it is weak.
  • No ranking table and no league position. A country sits in a distribution, and its confidence sits beside it.
  • No policy advice. The agendas say what to raise and what to measure first, computed from the data, and stop there.
  • No score from a language model. Panel estimates stay in their own field, are labeled as estimates, and never enter a published capability score.

Where it is weak

The benchmark is early, and these are its current failures.

Three of the nine capabilities correlate closely enough with income that the current indicators cannot separate the two. Several dimensions rest on thin evidence, and the Delphi layer has no reviewed panel run behind it yet. All of that is written down:

  • The limits record every place the data is wrong about the world, from docs/KNOWN-ARTEFACTS.md.
  • The diagnostics test the model against itself: wealth sensitivity, redundant indicators, dimensions that overlap.
  • The decision log holds each methodological choice, what it costs and what evidence would overturn it.
  • The challenge page is where you file an objection, a dataset for a gap or a documented delivery.

The code, data and documents are in the open repository on GitHub. You can trace every number to the series it came from.

Who built it

Envisioning researches how societies anticipate and act on change. The benchmark is part of that work.

Envisioning is a technology research institute and advisory. Brazil is the first field case, and it gets no special treatment in the model. It does have a layer of its own: the Brazilian layer carries the institution map and the subnational spread in Portuguese. The benchmark itself is published once, in English.

Envisioning is looking for institutional partners to sharpen the method, close gaps with comparable data, and connect a diagnosis to an intervention. Support says what backing the work looks like. Anything else goes to contact.

Where to start

Four kinds of reader, and the page each one should open first.

If you report on a country
Open its profile for its nine scores, then the limits before you quote a number. Confidence sits beside every score and belongs in the sentence with it.
If you work inside government
Start at your country's agenda. It names what to raise, what to measure first and what to hold, computed from the scores rather than recommended by us.
If you do research
The method explains the scoring, the diagnostics test the model against itself, the glossary defines every term this project invents, and the output directory is a Frictionless data package with a schema per published shape.
If you want to help
Ways to help lists every way to help, cheapest first. The fastest useful one is naming a dataset for an open gap.

Check the confidence before quoting a score. A thin dimension rests on one or two indicators, and the viewer marks it wherever it is drawn.