How the numbers are made

Sources (RSS, GDELT, official feeds, scraped outlets) are ingested continuously and routed to dynamics by topic. An extraction model proposes atomic claim candidates with quotes. A clustering pass folds near-duplicate reports into one lead and counts the fold as corroboration. An adversarial verification model recommends approve/review/reject with rationale — carrying each source's measured reliability and any state-affiliation label into that judgment. Then a human editor decides. The only exception: claims from tier-1 sources with a clean AI verify and confidence ≥ 0.8 may auto-publish under a policy the pipeline cannot widen — and sources that are state-controlled or below a reliability bar are excluded from even that.

Each dynamic scores 0–100 on four components, daily. In brief: Volatility — anomaly of event/claim arrival rates vs a 28-day baseline, plus live telemetry anomalies (e.g. AIS vessel counts). Uncertainty — verification reject rate, extractor confidence spread, forecaster dispersion, hedged language. Complexity — actor count, relationship variety, domain breadth. Ambiguity — contested-claim share, unresolved candidates, and measured worldview divergence between published factions. The headline is the component mean. Young dynamics blend against an analyst-seeded baseline; the real-signal weight rises automatically with history and is disclosed on each dossier's score breakdown. Confidence reflects source coverage, freshness, provenance, and agreement. Every daily snapshot stores its full input vector — the “why this score” module on each dossier renders those stored inputs directly.

Questions are binary, dated, and resolvable by a stranger from public sources — criteria are written before opening. Submissions are scored with Brier and log scores at resolution; the crowd aggregate excludes our AI baseline (@vuca_ai), which is scored as a benchmark and published win-or-lose. The full board is on /accuracy.

Sources are scored from their measured record in this pipeline only: P = approvals + 0.5·corroborations given + 0.5·corroborations received; N = rejections + 2·later contests + verifier flags; score = 100·(P + K/2)/(P + N + K) with K = 6 shrinking new sources toward a neutral 50. Duplicate folds are neutral. Identity and state-affiliation labels are machine-drafted with cited evidence and publish only after editor review. Full ledger: /sources.

A weighted blend — ambiguity 40%, worldview divergence 35%, contested-claim share 15%, forecaster disagreement 10% — renormalized over whichever inputs exist for a dynamic, so the index is honest from day one and sharpens as coverage deepens: /divergence.

Dossier dashboards (dollar security, PLA activity, the Ukraine Initiative Index) are built from named public series — FRED, IMF COFER, Taiwan MND bulletins, Oryx confirmed losses, DeepState map layers — each cited on the widget, with method notes where our computation adds a step (e.g. geodesic area used for momentum only). Structural series never feed the live VUCA score; they are context, and the engine excludes their namespaces explicitly.

Coverage is uneven across theaters; visually-confirmed loss data undercounts; assessed-control maps carry their makers' judgment; young dynamics lean on seeded baselines; and models make mistakes — which is why machine output is human-gated by default (the sole exception, the labeled tier-1 auto-publish lane, is policy-bounded and auditable), and why the record of our misses stays public.