Why the benchmark works this way

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

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

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 Do not edit a decision in place: supersede it with a new one and mark the old as superseded, so the reasoning stays auditable.

D1: Nine dimensions, no headline ranking

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

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

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

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

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

Superseded by D16 on 2026-08-26.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

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

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

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

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

D3: Equal weights inside a dimension

Choice. Every indicator in a dimension counts the same.

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

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

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

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

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

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

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

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

D5: Missing data is dropped, never imputed

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

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

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

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

D6: Dimensions are scored at any coverage above zero

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

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

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

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

D7: Winsorize with Tukey fences at k = 3

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

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

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

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

D8: Only the most recent observation, no trends

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

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

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

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

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

D9: Gap indicators stay in the registry

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

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

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

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

D10: Inspectability is a hard filter on sources

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

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

Cost. Real coverage loss in Coordination and Shared Purpose.

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

D11: Delphi output never enters the capability score

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

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

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

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

D12: Panel disagreement is recorded, not averaged away

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

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

Cost. No single clean number per cell.

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

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

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

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

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

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

D14: Provenance is stored, never inferred

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

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

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

Overturned by. Nothing.

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

Choice. One adapter. Everything else is a gap.

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

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

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

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

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

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-26.

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

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

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

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

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

D18: One display for every 0 to 100 score

Recorded 2026-08-26.

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

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

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

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

Cost.

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

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

D19: Extended countries get no visual marking

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

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

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

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

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

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

D20: Documented deliveries are recorded as evidence and never scored

Recorded 2026-08-26.

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

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

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

Recorded 2026-08-26.

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

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

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

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

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

What changed:

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

Cost.

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

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

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

Recorded 2026-08-26.

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

Two rules make the number mean something.

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-26.

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

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

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

Effect on the wealth correlation:

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

Cost. This is the important part.

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

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

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

Recorded 2026-08-26. Extends D22.

Choice. Three changes to the trend layer.

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-26.

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

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-26.

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-26.

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

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-26.

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

Three rules govern their use.

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

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

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

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

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

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

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

Recorded 2026-08-26.

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

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

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

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

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

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

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

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

D30: Every number opens onto the field it sits in

Recorded 2026-08-26.

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

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

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

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

Cost.

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

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

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

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

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

Recorded 2026-08-26. Extends D20.

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-26.

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

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

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

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

Cost.

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

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

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

Recorded 2026-08-27.

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

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

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

Cost. None beyond the three characters.

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

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

Recorded 2026-08-27.

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

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

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

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

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

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

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

Choice. Two additions to EvidenceRecord.

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

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

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

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

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

Cost.

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

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

D37: The output directory describes itself

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

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

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

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

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

Recorded 2026-08-27. Extends D35.

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

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

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

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

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

Cost.

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

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

D47: Every country sets the frame it is scored against

Supersedes D16. Recorded 2026-08-27.

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

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

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

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

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

Cost.

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

D51: Latin America is covered whole, not sampled

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

D54: An institution map explains capability and never scores it

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Recorded 2026-08-29. Implements issue 1.

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

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

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

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

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

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

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

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

Recorded 2026-08-29. Implements issue 2.

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

D65: Provisional layers need evidence before public promotion

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

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

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

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

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

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

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

D66: Subnational observations corroborate the national comparison layer

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

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

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

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

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

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