Published verbatim

Honesty rules (G1)

• No claim from a paper less than 14 days old may display a status above UNTESTED_*.
• REPLICATED requires at least 2 independent qualifying evidence items in the ledger.
• A language validator rejects any user-facing generation containing forbidden lexemes (proven, guaranteed, works, confirmed) for statuses below REPLICATED. Validator failures block publication. Zero tolerance in CI.
• A failed automated reproduction never auto-publishes as FAILED_REPLICATION; it routes to editor review.
• Every published verdict or report carries: status, confidence band, evidence trail with timestamps, coverage statement (for runs), and a changelog.
• Every private report is self-contained (coverage statement inseparable from the headline finding) and registered in the hash-verification endpoint.

Calibration commitment

Brier is computed from stored confidence_probability on resolved verdicts only, with a 24-month horizon. Unresolved verdicts are reported as coverage, not hidden. We do not publish a curve until those numbers exist. See the Calibration page for the live empty state.