If one AI writes the code and the same AI declares it correct, you do not have verification. You have a student grading their own exam and reporting a suspiciously steady stream of A grades.

Fusion's answer is structural: the verdict comes from a jury, the jury is drawn from rival vendors, and no juror knows whose work they are grading.

The family bias problem

Models have families, and families have taste. An evaluator model tends to prefer output that resembles what it would have produced itself: the same idioms, the same structure, the same style of hedging. Nobody tuned it to cheat; preference for the familiar is just what similarity looks like from the inside.

That bias is tolerable when the stakes are a chat reply. It is intolerable when the verdict decides whether code merges, because the whole value of the verdict is its independence from the author.

What cross-vendor buys

Seating judges from competing vendors, in the standard configuration one each from the Claude, Codex, and Gemini families, changes the failure math:

  • Correlated blind spots decorrelate. Where one family is habitually careless, another is habitually suspicious. A defect must slip past readers with genuinely different priors to reach a unanimous pass.
  • No brand holds the gavel. The verdict cannot be dismissed as one vendor's marketing grading its own product, which matters the day the receipt is challenged in a real dispute.
  • Style stops mattering. With mixed families on both sides of the table, "reads like my kind of code" stops being an advantage any candidate can bank on.

The blindfold does the heavy lifting

Cross-vendor seating alone is not enough, because judges could still play favorites if they could see names. So packets are blinded: identities stripped, candidates relabeled, ordering randomized independently per judge. The map from letters to authors is sealed until every vote is filed.

Blinding produces a pleasant surprise: a judge from the same vendor as a candidate is acceptable, because it cannot tell which candidate is its sibling. Neutrality lives in the process, not in the seating chart alone.

Votes, not vibes

Each judge files a criterion-level verdict against the frozen Definition of Good: satisfied or not, with reasoning, line by line. The majority carries; the dissent is preserved verbatim. Two passes happen in a full tournament: one to rank the candidates, one to grade the integrated final as a fresh piece of anonymous work.

The receipt then shows the vote as it happened. A unanimous pass and a 2–1 pass are both passes, but they carry different confidence, and hiding that difference would be a small lie in a product whose entire premise is not telling them.

The claim, precisely

We do not claim a cross-vendor jury is infallible; three readers can share a blind spot, and the receipt never says "perfect", it says "these criteria, checked this way, by these processes." The claim is narrower and more useful: the verdict was produced by a process with no structural reason to favor the author. In a market where most evaluation is a model admiring its own reflection, that is the difference worth paying for.

Source ledger

  • Merged strategy, sections 2 and 5.
  • Fusion MVP PRD, FR-12 through FR-16a.
  • Fusion planning model, section 6.