Cheap intelligence makes capability the bottleneck
Intelligence is getting easier to reach. What stays scarce is the ability to choose a direction, coordinate people and build what the technology makes possible. That ability is what this benchmark tries to measure.
What changed
Agents lower the cost of execution, AI speeds up science, and robotics carries both into physical production.
A capability that once took decades to build can now arrive in months, through tools and knowledge developed somewhere else. Knowing how a thing works matters less when more people can reach the means to make it.
The bottleneck moves to choosing a useful direction, coordinating the people who will act, trusting their work, deploying what they build, maintaining it and scaling it. Those are institutional and social capacities. The benchmark already measures them, and falling execution costs make them matter more.
Where the claim holds, and where it fails
The claim comes in a strong and a weaker version. Each sentence below is computed from the current release and changes when the data does.
Countries can be rich without being equally able to anticipate change, coordinate, learn, experiment, adapt, build or hold a shared purpose. The benchmark tests that in two strengths. The strong claim is that capability is separate from wealth, so the nine capabilities do not simply follow income. The weaker claim is that countries at the same income have different capability shapes, so a country's profile says something its income does not. If both fail, the benchmark is an income table with extra steps.
Are the nine really one thing?
If the strong claim is wrong, the nine capabilities collapse into one factor that tracks GDP per head. The diagnostics measure that directly: how much of the variation across countries one first factor carries, and how closely that factor follows income.
Over the 115 countries with all nine capabilities scored, one shared factor carries 49.4% of the variation. The same number of unrelated capabilities measured on 115 countries would give 16.1% by chance, and 17.7% in 19 draws out of 20. So the capabilities do move together, and 50.6% of the variation sits outside that one factor.
Across the 115 of those countries with an income figure, the shared factor correlates 0.81 with log GDP per head, so income accounts for 66% of it. The shared factor looks like income. What the benchmark can add beyond income sits in the other 50.6% of the variation and in how a country's capabilities differ from each other.
Each point is one dataset release, in order. The shaded band runs from the average share chance produces at that release's number of countries to its 95th percentile. A hollow point reads fewer than nine capabilities, because too few countries had all nine scored that release. The dashed line in the lower panel marks 0.8, at or above which the page reads the shared factor as income.
Across 33 dataset releases the share has run from 48.5% to 70.4%. The largest move between two releases came at dataset 7.0.0, from 60.4% on 33 countries to 49.8% on 33.
Capability by capability, six of the nine sit below the wealth-tracking line and three sit at or above it.
- Shared Purpose0.25118 countries
- Trust0.35123 countries
- Agency0.45123 countries
- Coordination0.53123 countries
- Adaptability0.53123 countries
- Building0.57123 countries
- Learning0.71123 countries
- Experimentation0.72120 countries
- Anticipation0.86123 countries
Absolute correlation with log GDP per capita, on a 0 to 1 axis. The line marks 0.7, the point at or above which the diagnostics call a dimension wealth tracking. Country counts differ because a dimension below the coverage floor publishes no score.
Shared Purpose moves least with income, at 0.25. Anticipation moves most, at 0.86, which is close enough to income that the project treats it as a known failure and says so beside the number.
What is left after income?
For every country and capability, the wealth residual is the gap between its score and the score its income predicts. Four tests ask what those leftovers hold, across the 115 countries with all nine. Each prints counts and averages over countries, and no country's leftover is published.
Take out of each capability the part its income predicts, and the leftovers still move together. Across the 115 countries with all nine, one factor carries 30.9% of what is left, where unrelated leftovers would give 16.2% by chance and 17.7% in 19 draws out of 20. It runs mostly through Anticipation, Learning, Trust, Coordination and Experimentation.
Countries at the same income do not share a shape. A country's shape sits about as far from its 10 nearest neighbors in income as from countries picked at random (a mean distance of 1.40 against 1.43), and the shapes line up along patterns that leftovers dealt out at random do not produce: one pattern carries 22.8% of them, against 19.4% at the random 95th percentile. Neighbors still differ less than leftovers dealt out at random would (1.40 against 1.78), because part of what is left is a level: a country above its income line on one capability tends to sit above it on others.
The order of the leftovers mostly holds between releases. Across 35 capability-by-release comparisons where an indicator change moved them, the rank correlation with the release before fell as low as 0.72 on Trust (n 36), and never below 0.50, where the page would say the leftovers read the indicators more than the countries. Dropping any one country moves an income line by at most 0.45 standard errors, and that country's own leftover by at most 0.11 of the spread around the line, so no single country carries a line.
For the typical country, income accounts for 32.8% of how far its profile sits from the average profile, and 24.5% on average across 115 countries, pulled down by countries the income line misses by more than the average profile does. Little of a capability profile is income.
The diagnostics print every figure with its null and the rule that reads it, and decision D138 records that the rules were fixed before the tests first ran.
What does that mean for the claim?
The strong claim, that capability is separate from wealth, does not hold for what the nine capabilities share: that shared part is mostly income.
The weaker claim holds on this data. Countries at the same income have different capability shapes, and the shapes are organized beyond chance, so a capability profile carries information that income does not. The order of the leftovers mostly survives indicator changes, so the shapes are not only a product of which indicators were chosen. The margin over chance is narrow, 22.8% against 19.4%, so this is the reading most likely to change as countries are added.
All of it rests on 115 countries with all nine capabilities scored. At that size a handful of countries can move any of these numbers, so each one is a hint that a wider country set could overturn.
The diagnostics also rescore the model with the wealth-correlated indicators removed. The limits record which dimensions lose every measured indicator after that removal. The method explains how a statistic becomes a score, and docs/WHY.md states the full argument and the evidence that would disprove it.
A correlation with income is not proof that income causes the capability. It marks a dimension whose current indicators cannot separate the two. That is a data problem, and the project publishes it.
Leverage: what multiplies a capability
The second layer covers the resources through which a capability gets multiplied.
Singapore and Brazil can reach the same frontier model. The systems around that model decide what happens next. A country with capability and no leverage leaves its ideas undeployed. The layer would measure the resources:
- AI access and compute
- data, connectivity and energy
- capital and technical talent
- robotics and advanced manufacturing
- scientific capacity and institutions that deploy emerging technologies
Velocity: how fast a capability moves
The third layer follows the rate at which a capability changes, because under fast change the level alone is misleading.
A country can hold a strong capability and still lose ground if its institutions cannot learn at the speed of the technology around them. The layer would measure:
- AI adoption and skill acquisition
- time from idea to prototype to deployment
- institutional experimentation and regulatory adaptation
- diffusion of new practices and business creation
- scientific adoption curves
Both layers are computed already, as offline fixtures the viewer does not publish. Neither reaches a score, a confidence or an agenda. Decision D65 sets what each one has to pass first: a settled method, a six-month review record and a review from outside the maintainers. Until then the foundation is the published layer and the other two are research.
Brazil is the first field case
One country tests whether the frame says anything a national statistics office does not already say.
Envisioning tests the thesis in a real institutional setting in Brazil, through its Brasil Capaz work on raising individual and collective capability. The benchmark gives that work a target: nine scores to raise and a set of indicators that says which parts are measured and which are guessed.
Brazil gets no special treatment in the model. It does have a layer of its own: the Brazilian layer carries the institution map, the subnational spread and the agenda in Portuguese, beside the English profile. The argument for the program is on that layer, in the language of its readers.
What the benchmark is for
It gives a public-sector leader, a journalist and a citizen one question and one body of evidence.
The benchmark exists so that different readers can ask what a country can actually do and check the answer against its sources. It sits inside Envisioning's wider work: signals detect exponential change, the benchmark measures whether a society can absorb it, and the capability programs and institutional design labs work on raising the number.
Read the decision record before quoting a result. This page states the argument. It does not replace the sources or the limits.