Confidence & consensus
Publication is decided by independent two-model agreement. These thresholds are the exact ones implemented in scripts/lib/consensus.mjs.
Publication bands
| Agreement | Meaning | Result | Confidence |
|---|---|---|---|
>90% agreement | both model families corroborate the clause and risk | published | normal |
70–90% agreement | both flag the surface but clause/risk differ | published | low |
<70% agreement | no independent corroboration (contested) | suppressed | — |
The verification gate (applied before consensus)
A finding is published only if its verbatim quote re-derives as an EXACT substring of the stored snapshot (offset-anchored, else exact substring; a stale offset never accepts the wrong text). Unanchorable quotes are rejected.
What this means for you
A published finding was independently corroborated by two model families and its quote re-derives exactly from the archived source. A low-confidence finding is published but flagged because the families agreed the surface is risky while differing on the precise clause or rating. A contested finding (no corroboration) is suppressed and never shown.
Generated from live pipeline data. Informational only, not legal advice.
AI Policy Intelligence Brief
Built for compliance officers, legal counsel, and SaaS founders. Subscribe to the email digest — one short brief when a tracked vendor materially changes its terms, training policy, or risk rating. Prefer in-app? Watch platforms in your alerts inbox instead.