Every score below is computed deterministically from recorded data. There are no AI-generated trust scores and no hidden calculations. Where data is insufficient, we say so rather than fabricate a number.
Each source's health score (0–100) is computed from observed capture telemetry: success rate ×0.6 + parse success ×0.2 + recency ×0.2, minus a penalty for consecutive failures. Bands: 95–100 Excellent · 80–94 Healthy · 60–79 At Risk · <60 Failing.
Data used: capture-event success/failure history and last-success age. Limit: a source with too few captures reports as unknown rather than a confident score. HSI observes sources and reports protection (Cloudflare, login walls); it never bypasses them.
Relationships between sources, signals, evidence, claims, reports, and corrections are built deterministically from real foreign keys and recorded tallies — never inferred by a model. Every edge carries provenance (the exact fact it came from) and a confidence: explicit links are 1.0; stance edges use the recorded support/challenge share.
Limit: relationships without a recorded link are not created — absence of an edge means absence of evidence, not a hidden judgement.
A narrative is a cluster of graph members expanded deterministically from a seed over real edges. Its stability score weights coverage, source diversity, evidence density, low contradiction, and size, minus failing-source penalties. Source diversity and actor diversity saturate (more distinct entities → closer to 1).
Data used: the evidence graph + source monitoring. Limit: narratives are human-seeded; the engine only expands and scores membership — it does not invent narratives, assign opinions, or score truth.
Forecasts are recorded before outcomes are known, then evaluated against reality. For probability forecasts we use the Brier score (squared error); a forecast is correct when its stated side matches the outcome. Calibration compares stated confidence against actual accuracy per confidence band; the calibration score is 1 − mean |predicted − actual| (1.0 = perfectly calibrated).
Confidence bands: 0.0–0.1 … 0.9–1.0. Data used: recorded predictions + recorded outcomes only. Limit: there are no AI predictions — humans/agents record the forecast; the engine only scores it. Unresolved and expired forecasts are reported as such, never counted as correct.
Operational agents automate analysis with fixed, deterministic rules over the systems above. Every agent run, action, and finding is recorded and reproducible. Agents never publish, review evidence, moderate, or change trust scores — they observe and report.
These metrics describe HSI's process and track record, not absolute truth. Counts reflect published/aggregate data only; unpublished items, private evidence, reviewer identities, and internal notes are never exposed here. When the underlying data is too thin to support a score, the portal shows "—" or 0 rather than a fabricated value.