The desk's rules, in public

Machines draft; editors decide. Claims, narratives, faction stances, analyst briefs, forecast questions, and source profiles are all machine-proposed and human-reviewed before publication. One disclosed exception: a claim may auto-publish when it comes from a tier-1 source, passes AI verification cleanly, and carries confidence ≥ 0.8 — a policy enforced server-side that pipeline code can only narrow, never widen. Auto-published claims are labeled in their provenance.

VERIFIED is reserved for primary-source facts. ASSESSED marks single-source press reporting. CONTESTED means credible evidence now cuts both ways — we show the dispute rather than pick a side quietly. Attribution claims and election-irregularity claims always require human review. State-controlled outlets are quotable for what a state says; their factual assertions about contested events need independent corroboration.

The record is append-only. We do not silently rewrite: when evidence changes, a claim is revised or contested with a logged revision, and score history is never recomputed. The full revision log — including the claims we got wrong — is part of /accuracy. If you believe something is wrong, tell us via /contact — the fix will be visible, not quiet.

No source pays for standing; the Source Ledger is computed from pipeline behavior and cannot be bought. No advocacy funding. Scores and assessments are analytic judgments, not financial or policy advice.

Extraction, verification recommendations, clustering, draft narratives, and the @vuca_ai forecast baseline are model-generated and labeled as such where they surface. Model outputs never assert probability as fact — “how likely” on this site always means a scored forecast with a public track record.