Skip to content

The tribunal

Judged, then audited, before you see it

Every draft goes through a judge and an independent auditor before a human reads it. The full transcript is stored with the article.

  1. 1

    DRAFT

    Sonnet-tier writer composes a long-form draft against the persistent site graph — real internal links, target entities, author block.

  2. 2

    JUDGE

    Opus/Fable scores six dimensions (intent, entities, EEAT, structure, links, schema) into a 0–100 composite and ranks weaknesses by user impact. Adversarial by design.

  3. 3

    AUDITOR

    Independently re-derives link integrity, schema validity, and entity coverage, then flags any unattributed numeric claim. Trusts nothing; confirms with proof.

  4. 4

    REVISE

    Weaknesses + audit flags feed the next draft. Up to 3 loops. An article ships only at ≥80 with zero broken links, valid schema, and no unproven claims.

The same tribunal runs from the UI, the MCP server, and CI. Model routing keeps it economical: Haiku extracts, Sonnet drafts, Opus/Fable judges & audits.

Method — the tribunal · ProofReef