Some numbers here are wrong

Places where the v0 model produces a number that is wrong about the world rather than informative about it. These are not bugs in the code: the pipeline is doing what it was told. They are failures of measurement, and anyone building on this needs to know them before quoting a score.

Evidence for each is either a diagnostic in data/out/diagnostics.json or the in-session panel run in data/delphi/in-session-round1.json, which scored 144 of the 468 country-dimension cells against the same evidence briefs.

Numbers here are from dataset 4.0.0 unless a figure names the run it comes from. Where a figure predates the current frame it says so, because the scale it was measured on is not the scale in use.

A1: Experimentation is not measured, it is inferred from patents

Severity: medium.

Four of eight indicators are observed. GEM early-stage entrepreneurial activity and fear of failure carry half the dimension, so it no longer rests on patents alone. Venture capital, regulatory sandboxes, university spinouts and business R&D share are still gaps. Mean confidence is 0.226, above only Coordination, Trust and Shared Purpose, and the GEM series covers 16 of the 52 countries, so 36 are scored on patents and trademarks alone. See D21.

Resident patents and resident trademarks per head measure formalised, completed, defensible invention, which is close to the opposite of the many-cheap- experiments construct the dimension is supposed to capture.

The panel and the indicators disagree here more than anywhere else in the model. The four largest departures across the 16 countries the panel covers:

The Netherlands' 28.0 is not a finding about Dutch innovation. It is the absence of venture-capital data. Uruguay scores 34.3, having legalised and regulated a national cannabis market and run a fintech sandbox. Argentina scores 21.5, having produced more technology firms of scale per head than anywhere else in the region.

Fix. A venture capital series is still missing. The only inspectable aggregate, the OECD SME and Entrepreneurship Financing scoreboard, covers 6 of the 16 original countries and omits Brazil, India, South Africa and Singapore. Business R&D share is the next best candidate, from UNESCO or OECD research and development statistics. Extending GEM coverage to the other 36 countries is the cheaper half.

A2: Per-capita normalisation flattens India

Severity: medium.

India scores 9.8 on Experimentation and 26.7 on Anticipation. Both come from dividing absolute counts by 1.4 billion people. The arithmetic is correct and the result is not informative on Experimentation: it says India files few patents per head, which was never the question. The panel put India at 40 there, a gap of 30.2 points, and at 25 on Anticipation, which the indicators now match within 1.7.

Fix. Per-capita is right for most indicators and wrong where capability is concentrated in institutions rather than spread across a population. Consider a per-capita and absolute-capacity pair for research and experimentation indicators, reported separately.

A3: Coordination and Trust remain weakly separable from wealth

Severity: high. The first source-backed Trust release is still thin.

The current figures are from dataset 4.4.0. Coordination publishes a score for 44 of 52 countries, based on two operational indicators. Trust publishes a score for 36 of 52 countries, based on the Joint EVS/WVS A165 social-trust measure and the 2019-frozen contract-enforcement measure. Trust's dimension correlation with log GDP per capita is 0.627 (Spearman 0.711, n=35), so the release is usable as a research baseline but not a clean claim of wealth-free trust. D23 retired the WGI perception composites and D44 retired homicide.

Both dimensions leaned on the Worldwide Governance Indicators, which are expert and firm perception composites that track income closely by construction. On the 16-country run, with those indicators in place, Coordination correlated with log GDP per capita at 0.90 and Trust at 0.88.

Drop every indicator correlating with log GDP per capita at |r| ≥ 0.7 and each dimension kept exactly one indicator. The table is from that same run, and the indicators in it are retired, so it cannot be recomputed on the current frame:

The pattern to read is the class column. Every indicator that failed the old test was class P and every indicator that survived was class C. The current Trust release adds a class P social measure from a named academic survey to the class C contract measure. The social measure alone correlates with log GDP per capita at 0.669 in the current alignment, so it is a watch item rather than a verdict. Coordination has a 2024 World Bank budget-execution series beside the older border-time measure, but that combination remains an operational proxy with uneven country coverage.

Fix. These dimensions need observable, behavioural indicators that are not WGI and not frozen at 2019. Budget execution is a useful API-backed addition for Coordination, but it does not show whether agencies delivered the same objective. Trust needs pooled EVS/WVS rows, a recent court-throughput or case clearance series, and broader institutional-performance evidence. See also A9, which is the same problem seen from a small competent state, and D20, where documented cross-agency delivery is being collected as evidence.

Overturned by. Behavioural indicators that cover the country set and show that each dimension remains distinct from income after its indicators are combined. Until then, treat both as thin operational proxies and treat the Trust score as provisional.

A4: The four WGI series are one measurement wearing four names

Severity: closed. None of the four are scored. See D23.

Government effectiveness, regulatory quality, rule of law and control of corruption correlate with each other between 0.93 and 0.98 on the ten-country run that last scored them. They are spread across Coordination, Trust and Shared Purpose, so a single underlying perception measure is being counted three times in three different dimensions.

None of the four are scored. The measurement-class analysis in D23 argued for keeping none of them: they were the mechanism turning three dimensions into restatements of income per head. The rows stay in the registry as retired, with the reason on each.

A5: Voice and accountability is answering a different question in Shared Purpose

Severity: closed. The indicator is retired. See D23.

GOV_WGI_VA.EST measures the democratic channel for participation. Shared Purpose asks whether people can see themselves in a common project. Singapore scores 20.9 while being one of the most effective collective actors in the set. The panel put it at 55, splitting the difference between very high collective-action capacity and genuinely limited pluralism.

The spec is explicit that political uniformity is not a capability, so the fix is not simply to raise Singapore.

Voice and accountability is retired, so Singapore's 20.9 is gone. Shared Purpose now rests on two indicators, tax revenue and income inequality, at mean confidence 0.223, and it publishes a score for 46 of 52 countries. Its correlation with log GDP per capita is 0.45, the lowest of the seven dimensions that publish. The behavioural measures it needs are civic participation and volunteering, both still gaps, plus voter turnout, which is published by International IDEA and would have to be entered by hand.

A6: Doing Business indicators are frozen at 2019

Severity: low, correctly handled.

Five indicators come from the discontinued Doing Business programme: time and procedures to start a business, border compliance time, contract enforcement time, electricity connection speed. All are stuck at 2019 and the World Bank has archived the codes.

The recency term already marks them down, so this is visible rather than hidden. It is listed because these five have to migrate to B-READY, which is in the API as IC.BRE.* inside World Development Indicators and covers 12 of the 52 countries in its 2024 round. The migration waits on coverage and not on the publisher. See A12.

A7: Learning overstates Brazil and understates Korea, Estonia and Singapore

Severity: medium.

The Learning dimension leans on enrolment and expenditure, which are inputs, plus the Human Capital Index, whose last full round is 2020. Countries with high enrolment and weak outcomes score well. Countries with exceptional measured outcomes do not get credit for them.

Against the panel, Korea is 29.1 points low, Singapore 29.0 and Estonia 28.2. The error runs the other way where enrolment rose fast and measured outcomes did not follow: Peru is 19.7 points high, Mexico 14.8, Colombia 12.3 and Brazil 8.2.

Fix. A learning-outcomes series (PISA or PIAAC) would resolve most of this. It is a gap because coverage across the country set is uneven, not because the data does not exist.

A8: Every correlation here is a hint, not a result

Severity: structural.

Every correlation in diagnostics.json is computed on the 52 countries loaded. Fifty-two points is enough to reverse a finding and not enough to establish one. Two dimension pairs sat at 0.94 on the 16-country run and read as near-duplicates. At 52 no dimension pair passes the redundancy threshold at all, and the highest is Anticipation with Agency at 0.83. The nine dimensions separate when the sample is wide enough to separate them.

The redundancy and wealth-proxy findings are strong enough to act on because they also have a mechanical explanation, not because the coefficient is large.

Do not report any of these correlations as established until the country set is substantially larger, and never quote one without its n.

A9: Coordination reads far too low for small, competent states

Severity: medium. No score carries it today, and it will return with the first replacement indicator.

Coordination now publishes a thin score for most countries, but the artefact is still active because the measurement error is a property of the indicator type, not of one run.

On the perception layer D23 retired, and on the frame in use then, Uruguay scored 18.8 on Coordination. In the same decade it delivered Plan Ceibal, a national digital government stack, and the first nationally regulated cannabis market in the world, each of which required several institutions to move together and hold position for years. Costa Rica showed the same pattern more mildly at 33.7, against a national choice in 1949 to abolish the army and redirect the money into health and education that has now held for three generations.

Those scores came from the Worldwide Governance Indicators and the Logistics Performance Index. Both are surveys weighted toward how a country looks to international business, and a small country with a small port does not look like much through that lens whatever its state can actually organise.

Fix. The same fix as A3. Coordination needs observable evidence of institutions acting together, not perception surveys of business conditions. Delivery records for cross-agency national programmes would be the direct measure, and they are a declared gap.

A10: The frame is 52 countries wide, and they are not the world

Severity: structural.

Every country in the benchmark sets the endpoints of every indicator scale and is measured against the result. See D47. So 0 and 100 mean "weakest and strongest of these 52", and the 52 were picked to expose contrasts and to cover Latin America whole (D51), not to sample the world. A score is a position in this set and carries no claim about a country outside it.

Two consequences follow.

Scores are only comparable inside one dataset version. Adding a country moves the endpoints it touches and restates every number. That is done as an announced rebase with a major version bump, and 3.0.0 numbers do not compare with 4.0.0 ones. Anything quoting a score has to quote the version with it.

Clamping has moved to history. No observed cell clamps: 0 of 1,604, because a current value cannot fall outside a frame its own country helped build. The outOfFrame flag now fires only where a historical value sits outside the current frame, which is 64 of 603 momentum baskets. A trend carrying a clamped basket member is part distance-to-the-clamp rather than movement in the country, and every surface that prints a trend prints that count.

A11: Building measures industrial output, and reads as delivery capacity

Severity: medium.

Building asks whether a country can build and deliver. Its five measured indicators are manufacturing value added, high-technology export share, labour productivity, electricity connection speed and economic complexity. All five describe industrial output. Nothing in the measured set can see a national programme that was specified, funded and delivered.

Brazil scores 25.3 at confidence 0.544, one of the better evidenced of its nine dimensions. The score is a correct statement about Brazilian industrial output and it is read as a statement about Brazilian delivery capacity, which is a different construct. In the same decade Brazil built and ran Pix, which settled 7.98 billion transactions in July 2026, and GOV.BR, which reports 175 million active accounts.

The two indicators that would carry the delivery construct, large_project_delivery and firm_scale_up_rate, are both gaps.

Fix. Two parts, one done and one open. Documented deliveries are now recorded in data/evidence/records.json against the gap they bear on, outside the score, so the cases are written down with sources and limits instead of being argued in prose. Thirty-three records cover 19 countries and bear on 10 different gaps. Brazil's 11 run from Embrapa in 1973 to Pix in 2020. See D20. The open part is a comparable delivery series across the country set, without which the gap cannot be promoted to an indicator.

Watch for. The same reading error in reverse. A country with strong industrial output and a poor record of finishing public programmes scores well here, and the benchmark currently has no way to say so.

A12: Coordination and Trust are scored on thin evidence

Severity: high. Trust is partly measured and Coordination remains narrow.

The figures in this entry are from dataset 4.4.0.

Coordination's new score is visible with its confidence and its two indicator rows. It remains a narrow operational proxy: budget alignment and border processing do not show whether several institutions delivered a shared national objective. Trust now prints a provisional score where both rows are observed, but every published score rests on exactly two indicators at confidence 0.159. Sixteen countries remain below the coverage floor. Shared Purpose sits one indicator above the floor and prints, drawn dashed with a marked axis and a confidence band that says do not quote it alone. That is a mitigation and not a fix.

Trust still has a narrow family balance. D57 splits the dimension into a social family, which asks whether people rely on strangers, and an institutional family, which asks whether they rely on courts, government and the civil service. The current social family has one observed row in 36 countries. The institutional family has one observed row in 51, while court performance and institutional trust remain gaps. familyBalance publishes this coverage, so a reader can see that the structural test is met only by the social plus stale contract pair.

Fix. Pool the held EVS/WVS country rows with respondent-level weights when the license permits, then land court throughput and case clearance, civil society strength, cross-agency delivery records, institutional trust, and behavioural measures of corruption experience. The generative panel can interpret the dimensions while those data are missing, but its values stay beside the indicator score and never become observations.

What the World Bank can and cannot supply. GF.XPD.BUDG.ZS, primary government expenditure as a proportion of the original approved budget, covers 44 of 52 at 2024 and correlates with log GDP per capita at 0.285. The value is two-sided: both underspending and overspending can indicate weak execution, so the registry converts it to absolute distance from 100 before scoring. The CPIA cluster covers only 10 of 52. Of the two Enterprise Survey corruption series, IC.FRM.CORR.ZS asks a firm what it believes firms similar to itself pay, so it records belief and is ineligible, while IC.FRM.BRIB.ZS asks whether the responding firm was itself asked for a bribe across six public transactions. The second covers 49 of 52, 44 of them at 2023 or later, and D60 wires it as a behavioural check: fetched, published beside Trust and excluded from the frame, the mean, the coverage floor and the confidence. It is not admissible as the score, because it puts the two-indicator dimension at about 0.53 against log GDP per capita where contract enforcement days alone sits at about 0.14, a larger wealth contribution than the one D44 retired an indicator over. Trust now has a source-backed score from the Joint EVS/WVS adapter, but still needs court data and a broader social and institutional comparison. The current shortlist is V-Dem, OECD Government at a Glance, and a harmonized court or audit source.

B-READY is what the frozen rows become. IC.BRE.* replaces Doing Business inside World Development Indicators, and its dispute resolution score is the most income-independent series in the sweep at -0.034. The 2024 round covers 12 of the 52 countries, so nothing is wireable from it yet. It is what Coordination's time to export and Trust's contract enforcement days turn into when its coverage arrives, which is the closing condition for A6 as well.

Civil society strength, voter turnout, volunteering and civic participation are absent from the catalogue under any database id, and the interpersonal and institutional trust items were never World Bank series. The Joint EVS/WVS adapter now supplies the first social row; Trust still needs an adapter for court clearance and Coordination needs one for anything past budget execution. The shortlist is ILOSTAT, IDEA voter turnout, and OECD government at a glance for court clearance. ingest: 'manual' remains available for sources with no usable API.

The one indicator added to fix contamination was the contamination: D42 showed that homicide_rate raised Trust's wealth correlation by 0.288, so D44 retired it. Generative estimates are therefore a useful research layer, but they cannot repair a missing observation series.

This artefact remains open while the dimensions rely on narrow proxies. For Trust, D57 sets the stricter condition: one harmonised social measure and one comparable institutional-performance measure, because two indicators from the same family would clear the floor without answering the dimension. The current release meets that structure with A165 and contract enforcement, but court performance, broader coverage and the wealth and redundancy review remain open. The budget series opens the Coordination door; it does not close its measurement problem.