Critic Score

A critic score is a 0-to-1 numerical rating assigned to each piece of pipeline-authored content by a separate reviewer model before publication. The score evaluates the draft against a structured rubric — factual accuracy against the underlying fact contracts, voice consistency with the publication’s editorial voice, internal coherence, and absence of the “AI tells” the publication’s voice constitution explicitly bans.

Each artifact type (executive summary, dashboard patch, section rewrite) has a configured critic floor. Drafts scoring above the floor publish automatically; drafts below the floor are held in a queue for analyst review. The floor for executive summaries is currently 0.65; a one-week soak period after recent prompt iteration determines whether to ratchet it up.

The critic is not a quality oracle. It is a defense against specific failure modes: drafts that pass the critic still get spot-checked by the analyst on Tier C surfaces, and the publication’s force-publish operator override exists for cases where the critic blocks legitimate prose. The critic’s value is in catching the bad cases, not in certifying the good ones.