A judge who knows the author is grading the author. Fusion's jury never gets the chance.
What a judge receives
When candidates finish, the run lead prepares one packet per judge. Each packet contains the frozen Definition of Good, the candidate diffs and evidence, and nothing else. Before the packet ships:
- every vendor name, model name, and branch name is stripped;
- candidates are relabeled with neutral letters;
- the ordering is randomized independently for each judge, so judges cannot even agree on which candidate is "A".
The identity map exists in exactly one place, held by the run lead, and it is only unsealed after all verdicts are in.
What a judge returns
Not a thumbs-up. A judge files a criterion-level verdict: for each line of the contract, satisfied or not, with reasoning. Then a recommendation for the best path forward, which the judge must give even if no candidate fully passes. "None of these, but B's approach to the migration is the one to salvage" is a legitimate and useful verdict.
Why identity-stripping is not paranoia
Models carry recognizable styles, and evaluator models carry measurable preferences. A judge that recognizes its own family's output faces a conflict of interest even with honest intentions; the politeness of machines is that they exhibit bias without ever feeling it. Blinding removes the opportunity rather than trusting the temperament.
The same logic allows something that surprises people: a model from the same vendor as a candidate can sit on the jury, because it cannot know which candidate is its sibling. Neutrality comes from the blindfold, not from the roster.
Disagreement is data
Judges vote independently and their votes are preserved as filed, including splits and dissents. A unanimous verdict and a 2–1 verdict both produce a decision, but they mean different things, and the receipt records which one you got.
Source ledger
- Merged strategy, section 5.
- Fusion planning model, section 6.
- Fusion MVP PRD, FR-12 through FR-16a.