How confidence works
Every claim gets a 0–100 confidence score built from five weighted components, minus penalties. The interface always shows the label and an explanation — never a bare number.
Score components
| Component | Weight | What it captures |
|---|---|---|
| Evidence directness | 30% | Primary measurement or document beats second-hand reporting. |
| Source reliability | 25% | Track record and institutional role of the source. |
| Independent corroboration | 20% | Genuinely independent confirmations — syndicated copies of one agency release do not count. |
| Recency and temporal fit | 15% | Fresh, time-consistent evidence scores higher than stale data. |
| Geographic / entity specificity | 10% | Precisely located, clearly attributed claims score higher. |
Penalties are subtracted for contradiction by observation, circular reporting, stale data, unresolved identity and single-interested-party sourcing. Each penalty is recorded with a code, points and reason, and shown on the signal page.
Labels
| Label | Range | Meaning |
|---|---|---|
| Verified | 90–100 | Direct evidence or source data, domain rules satisfied. |
| High confidence | 75–89 | Strong sourcing with independent corroboration. |
| Medium confidence | 55–74 | Credible but incomplete corroboration. |
| Low confidence | 35–54 | Limited or partly conflicting evidence. |
| Unverified | 0–34 | Single or weak sourcing; treat as a lead, not a fact. |
Disputed is a separate flag independent of the numeric score: a claim can be high-confidence and still disputed by a named party.
Source independence
The corroboration model detects shared press releases, single-agency syndication, same-ownership outlets, parties to a dispute and analyses built on the same dataset. These are collapsed into one effective source before corroboration is counted.