Scoring-model changelog

Public, dated record of every change to how NationsHelm computes scores. Versioned independently of the app (lib/scoringModel.ts → shown on /methodology). MAJOR = reorders headline rankings; MINOR = new/removed metric or floor; PATCH = fix/label.

2.10.0 — 2026-07-23 — Bloc taxon opened: type-derived, not a hand-picked allowlist

The Geopolitics "blocs" signal no longer scores a hand-picked list of ~30 orgs. That list was arbitrary, silently incomplete (it omitted Council of Europe / OSCE / WTO, so Ukraine read "No bloc memberships on record" despite deep Euro-Atlantic integration), and its gaps fed the 2.9.0 fake-zero. The taxon is now open and derived from the source:

• Ingest (scripts/fetch-country-blocs.mjs): take all current Wikidata P463 memberships for every nation (deduped by entity), not a 30-org allowlist. globalInfluence.blocs becomes the nation's memberships that are geopolitical/economic alignment blocs; globalInfluence.blocMembershipsTotal records the full count.
• What counts as a bloc — by the org's own type, not by hand. An org is scored if its Wikidata type (P31) is an alignment type (military alliance, regional organization, trade bloc, customs/economic/monetary union, supranational union, geopolitical group, cartel, …) within a size band (≥4 members — drops bilateral/trilateral vanity arrangements; ≤120 — universal bodies carry no signal), plus a small explicit supplement for significant blocs Wikidata under-types (OSCE, Council of Europe, SCO, OECD, OAS, OIC, Francophonie, Pacific Islands Forum). This yields 35 scored blocs; universal UN-system bodies (UN/WHO/WTO/ITU…) and technical/functional IGOs (development banks, hydrographic org, treaty regimes) are shown in the total but not scored — they don't discriminate alignment. (data/bloc-registry.json records every org, its member count, and its scored flag.)
• Why not raw count or IDF. Empirically: raw P463 count saturates (every state is in dozens of UN-system bodies); IDF/rarity-weighting ranks backwards (a 4-member Five Eyes outweighs NATO/EU); a pure size band admits technical bodies (Germany scored 44 on dev-banks + hydro org). Type + size + supplement is the combination that isolates alignment blocs.
• Score (lib/blocs.ts): the count of scored blocs on a diminishing curve, 100·(1−e^(−n/4)) — never saturates (n=1→22, 4→63, 9→90), replacing the old count×20 (which capped at 5 blocs). Null only when a nation has no membership data at all (blocMembershipsTotal === 0, a territory Wikidata doesn't cover) — renormalised out. A nation in memberships but no alignment bloc (North Korea: UN-system only) scores a real 0.
• Impact. Bloc-driven Geopolitics re-scores across the cohort with sensible ranks: Germany 86, France 89, US 83, Russia 83, Switzerland 71, Ukraine 63 (GUAM/OSCE/CoE/CIS-FTA — was 13/"blocless"), China 53 (SCO/BRICS/G20), North Korea 0 (was 13), Vatican 0. Micronesia/ Marshall Is./Palau now carry the Pacific Islands Forum (a Wikidata P463 gap, still hand-patched). The Global Influence card shows scored blocs + a total; the "No bloc memberships on record" copy is retired. Overall headline rankings are unaffected (blocs → Geopolitics → Strategic Position, not Overall). Snapshot diff: docs/data-quality-fix/snapshots/<published-id>-diff.md.

Supersedes 2.9.0 (the fake-zero null-fix): the null-vs-0 handling is now expressed through blocMembershipsTotal, and the empty-list ambiguity is gone because the taxon is comprehensive.

2.9.0 — 2026-07-23 — Bloc membership is null, never a weighted zero

An empty tracked-bloc list is no longer scored as a real 0 inside the Strategic Position Geopolitics component (lib/strategicPosition.ts) or the Global Influence metric-category (lib/insights/metricCategories.ts). Both now route the bloc signal through one blocMembershipScore(blocs) helper that returns null for an empty list, and renormalise over the signals that are actually present:

• The fake-zero. blocsScore = blocs.length × 20 turned an empty list into a hard 0 that entered Geopolitics at 40% weight. Because [] is a structurally-valid value (not null), it slipped past every missing-data guard — the AGENTS.md "missing is null, never a fabricated fallback" rule was violated in silence. Ukraine (in no tracked exclusive bloc — its Council of Europe / OSCE / WTO memberships are out of the tracked scope) scored Geopolitics 13, dragging Strategic Position to 46.
• The fix. Empty → null → renormalised out. Geopolitics still requires ≥1 sourced quantitative signal (soft power or GDP share) — a bloc list alone can't assert a score (Taiwan: no formal bloc, real weight), so the "no quantitative signal → null" guard is unchanged. The weight ladder (0.4/0.3/0.3, with the 0.57/0.43 two-signal fallbacks) is kept verbatim — only a null signal now drops out — so every present-bloc nation is byte-identical to before and the diff is confined to the empty-bloc set.
• Impact. The 22 nations with an empty tracked-bloc list get an honest, renormalised Geopolitics (or "No data") instead of a laundered wrong number: Ukraine Geopolitics 13→22, Strategic Position 46→48; similar lifts for Georgia, Serbia, Bosnia, Taiwan, North Korea, and the microstates/territories. A silent-wrong composite becomes an honest one.
• Coverage repair (data). Micronesia, Marshall Is., and Palau — full Pacific Islands Forum members the Wikidata P463 fetch silently dropped and the manual patch missed — are added to MANUAL_ADDITIONS (scripts/fetch-country-blocs.mjs) and backfilled in data/nation-profile.json; they now carry PIF instead of "No bloc memberships on record".

Not changed (documented follow-ups): the country-blocs ingest contract still measures an aggregate coverage floor and excludes zero-bloc nations from its payload, so it remains blind to a single nation dropping to zero — a per-entity/known-member guard is the recurrence fix. The tracked-org set (30 exclusive blocs) still excludes integration bodies (Council of Europe, OSCE, WTO); "No bloc memberships on record" copy is scope-accurate only for exclusive blocs. Snapshot diff: docs/data-quality-fix/snapshots/<published-id>-diff.md (0 blocking / 0 advisory — no Overall or Risk-board movement; the 14 empty-bloc nations with a quantitative signal get a renormalised Geopolitics, e.g. Ukraine 13→21 / Strategic Position 46→48).

2.8.0 — 2026-07-23 — Violence & Safety re-anchored to UCDP conflict thresholds

The Stability Violence & Safety sub-signal no longer inverts per-capita deaths against the worst country in the cohort. It is now min(perCapita, bandCap) (scripts/lib/safetyScore.mjs):

• perCapita — deaths per million on a log curve to a FIXED anchor (50 dpm → 0). Government Stability is state-continuity, which is relative to population, so 623 organized-violence deaths in a 1.4B state (an insurgency) is not scored as a war, while the same count in a micro-state is.
• bandCap — UCDP's OWN intensity thresholds cap the score so an absolute armed conflict can't read as "safe" even in a huge country (and 1 death ≠ 0): 0 → 100 · 1–24 → ≤90 · 25–999 (UCDP armed conflict) → ≤60 · 1000+ (UCDP war) → ≤15. Thresholds inherited, not invented.
• Fixed anchors, not a cohort max: a country's safety no longer moves because an unrelated country's war changed the maximum.
• Worse-of the two readings: the blend takes the lower of the finalized-annual (GED) and current-year (CED) safety, so a calm year-to-date partial can no longer launder over a country that was at war last year (the Iran "96 / Low political violence" bug → now "Active conflict", 15).

Impact: stability drops for conflict-affected nations (Iran 44→31, Syria 34→20, Russia 48→34, Mexico 53→39, CAR 45→23…); nations with zero organized-violence deaths are unchanged (safety 100). Overall headline board is stable (only RK 36–40 shuffle ≤4); Risk-top-15 gains Togo and Niger. The QoL active-conflict cap (D6) is unaffected — it still reads the raw per-million counts. Snapshot diff: docs/data-quality-fix/snapshots/snap-20260723-*-diff.md.

2.7.0 — 2026-07-22 — Diplomatic Conduct (leader Diplomacy dimension)

A leader's Diplomacy dimension (dip) now overlays from Diplomatic Conduct, a performance-under-test model (docs/diplomacy-conduct-v1.md) replacing the volume-dominated Diplomatic Signal. Two axes like Crisis Response: Test Severity (context — hostility of the external environment, from UCDP active-conflict intensity + conflictual event share + sanctions received) and Conduct (the scored blend), with severity amplifying the conduct deviation (k=0.4) so strong conduct under fire is worth more.

• Conduct components (renormalized over present weight, null below half): cooperative ratio 20% · restraint 20% (inverse of NET material force projected — QuadClass-4 initiated minus received, size-normalized — the leader-as-actor signal) · coalition breadth 15% · agreements & mediation 15% (CAMEO 045/046/057/036 as initiator) · sanctions balance 10% (imposed − received) · engagement 10% · leader foreign-media tone 10%. Directed GDELT source via scripts/generate-gdelt-conduct-bq-query.mjs.
• Why it changed: the old signal was 40% engagement volume + 30% cooperative share, so it measured diplomatic capacity (big active powers topped out) not conduct, and a belligerent leader projecting force abroad scored high. Restraint (net force initiated) is the fix; NET (initiated − received) separates an invaded defender who fires back from an unprovoked aggressor.
• Impact: leader Diplomacy re-scored across the active cohort. Zelenskyy holds (86, restraint 89 — invaded); Trump 90→75, Putin 76→58 (with traits), Netanyahu 81→59 (restraint 3). Diplomatic Conduct is a proxy — max confidence "medium".
• Known limits (documented): a symmetric war dilutes the aggressor's net (Russia's counter- attacked footprint), and proxy warfare isn't attributed to its sponsor (Iran) — GDELT records events, not who started a war. diplomaticSignal is retained on the record but deprecated.

2.6.0 — 2026-07-19 — D6: Quality-of-Life active-conflict hard cap

Quality of Life is hard-capped at 40 for any nation that is an active organized-violence party, correcting a structural blind spot: the Safety group's conflict metrics are population-scaled (deaths per million), so a large nation at war reads "safe" (Iran's armed- conflict deaths scored 93; Mexico's an average that didn't tank QoL), and where the primary homicide series is absent the Safety group nulls out entirely, leaving QoL as Healthcare + Education only (Iran QoL 80). Neither reflects that a war zone is not a high-quality-of-life place.

• Trigger (isActiveConflictParty, lib/indicators/scoringPolicy.ts): still active this year — UCDP-CED current-year deaths per million > 0 — AND at meaningful annual scale — UCDP-GED finalized deaths per million ≥ 10. Both readings are freshness-gated, so a resolved conflict cannot trip it. Requiring both keeps genuine wars (Ukraine, Russia, Iran, Israel, Syria, Mexico, Colombia, Myanmar, Pakistan, Lebanon, Cameroon) while sparing one-off localized violence a bare "any death this year" rule over-caps (Brazil 20 deaths, Thailand 7, Guatemala 1).
• Cap is a min(), not an average — high Healthcare/Education cannot renormalise it away. The category carries conflictCapped + cappedFrom so the UI shows "80 → 40, capped: active conflict" rather than an unexplained number.
• Impact (ranking diff): 11 nations' QoL drops to 40; Overalls fall 2–10 (Russia 58→48, Iran 49→41, Mexico 51→45, Israel 64→60, Ukraine 49→46, …). Behind the qolActiveConflictCap semantics flag (Group D), shipped ON after the diff review.
• Transparency (metadata-only republish, 0 value moves): conflictCapped + cappedFrom are now threaded through the snapshot projection (lib/snapshot/build.ts) and rendered on the category card as a "capped — active conflict" chip with the pre-cap number ("80 → 40"), so the score never reads as an unexplained contradiction of the visible group scores. The two UCDP conflict metrics' registry entries document the cap; the methodology page gains a "Why is a country's Quality of Life capped when it's at war?" entry.
• Not in this release: a matching active-conflict diplomacy penalty (applyConflictPenalty, scripts/lib/diplomaticSignalScore.mjs) is implemented in code but NOT yet activated in data — it is being reworked so leader-level violence is captured even when a war is fought off the leader's own soil (location-based UCDP misses it). Will land in a later version once settled.

2.5.0 — 2026-07-16 — Group D metric-semantics: net debt (D2) + unemployment gate (D3)

Two Group-D semantics options, signed off after review and shipped ON (the flags in lib/indicators/semanticsFlags.ts were inert stubs — this release actually implements them in buildCategories). D1 (public-spending band-optimum) stays OFF — a contestable normative band, not an error correction.

• D2 — Debt/GDP reads NET where the IMF publishes it. debtToGdp now scores from IMF general-government net debt (data/imf-net-debt.json, series GGXWDN_G01_GDP_PT, 2024, 84 nations) where available, else keeps the World Bank gross reading. Net uses a domain sharing the gross 150%-of-GDP worst-case ceiling but floored at 0, so net-creditor states clamp to 100 (Norway net −149% → 100; Australia gross 58 → net 32; Brazil 82 → 61). Nations the IMF does not net (Singapore — sovereign assets in GIC/Temasek) keep gross with the "Debt / GDP" label; net-scored nations render "Net debt / GDP". Flat-mean Fiscal Health (3 metrics) already caps any single reading's weight. 25 nations swap gross→net, 59 gain net where they had no gross, 43 keep gross.
• D3 — Unemployment gated on labour-productivity presence. In high-informality / subsistence economies a near-zero measured unemployment (Solomon Islands 1.44% → 100) is an artifact; where labour productivity is not sourced, unemployment is dropped and renormalised out of Labor Market rather than inflating it. Developed economies (productivity present) are unchanged. Cuba Economy 41→29, Solomon Is 60→58, Samoa/Tonga/PNG −2.
• Headline effect: 76 Economy movers, 34 Overall movers (all ±1–3). Top of the ranking reshuffles within the near-tied leaders — Denmark 85→86 takes #1 from Norway; Luxembourg 81→82 and Ireland/Estonia rise into the top-15; Japan 81→80 (its high net debt now counted). Momentum is a separate historical pipeline and is unaffected (as with Group A). Full before/after in ranking-diff-report.md; rationale + per-nation evidence in metric-semantics-decision.md.

2.4.1 — 2026-07-16 — Diff-review revisions

• A1 core-category rule — the ≥4 present categories must now include Economy AND Quality of Life. Nulls North Korea's Overall (no sourced Economy; it was building one from governance + QoL + innovation + stability); keeps the Pacific microstates that have real Economy + QoL.
• A3 War-null fix — a state with no measurable military (derived militaryScore = 0: no-army microstates and force-count data gaps) no longer scores the War scenario 0. War is null (renormalised out) below a measurable-military floor — absence is not "certain defeat". Saint Vincent fell from the #1 riskiest nation to #5; the homepage Risk Signal top-3 is now Yemen/Haiti/Syria. Every conflict state scores militaryScore ≥ 8 and is unaffected.

2.4.0 — 2026-07-16 — Data-quality remediation (Groups A–F)

Aggregation integrity (A)

• A1 An Overall now requires ≥4 of the 5 categories present (was: renormalise over any present set). Solomon Islands 75→50; Tonga/Grenada/San Marino corrected; nothing lost its Overall net (A4/A5/A6 restored the categories that should score).
• A2 Per-category coverage floors: ≥50% of a category's groups present and ≥2 sourced metrics. Worst-of Safety with no sourced homicide asserts only a real conflict threat, never a false-100.
• A3 Risk/Resilience gated on the nation carrying an Overall — data-sparse Gabon/Malawi/Libya/Hong Kong excluded from Risk; conflict states top the list.
• A4 Stability category = the 5-signal Government Stability blend (weighted registry groups), replacing the lone WGI Political-Violence series.
• A5 Innovation category now scores the Oxford GARI AI-readiness index (195 nations).
• A6 Governance coverage measured over groups, not the sparse raw-metric count — full-WGI Switzerland/Qatar/San Marino now score Governance.

Formula & consistency (B)

• B1 Difficulty-Adjusted score reconciled onto the single continuous multiplier (×1.00 at nation overall 70, clamped [0.92, 1.18]); methodology + registry + tooltip now agree; the phantom "Rating × meta.lck" definition retired.
• B2 Same-page contradictions fixed (water-scarcity on null, Peak Strengths reading the displayed stats, election-type vs free-fair-index labelling).

Coverage (C) — Taiwan V-Dem/V-Party join repaired via the canonical crosswalk (data lands on next ingest); dead GDELT media modules collapse and the false "~173 tracked hourly" claim retracted; leaked military honorifics stripped from leader names.

Infrastructure (F) — declarative registry fields (weight/status/band); canonical entity crosswalk + audit; leader-page ISR vintage aligned; data-quality CI gate.

Deferred (documented, not shipped): D1–D3 metric-semantics flags (OFF, pending sign-off); F3 ingest contracts / F4 versioned snapshot / F6 anomaly detection (designed).

≤ 2.3.x — pre-remediation

Continuous difficulty multiplier (was tiered); FSI dropped from Government Stability; frozen-indicator gating; momentum rebuilt on real history; RNG purged. See git history.