A Fusion tournament has two creative moments: invention, when candidates build rival solutions, and synthesis, when integrators combine the strongest parts into one final. Each moment gets its own blind jury pass, because integration changes the artifact. A verdict earned by the ingredients does not transfer to the mixture.

Pass one compares complete candidate approaches

Every candidate builds from the same frozen RunSpec and runs through the same protected gate. The first blind jury then compares only gate-qualified artifacts.

Its job is diagnostic as well as selective. A useful pass-one verdict identifies:

  • which candidate best satisfies each criterion;
  • which risks are unique to one approach;
  • which implementation choices should survive synthesis;
  • which regressions an integrator must avoid.

Those findings become shared input to the integrator stage. The integrators do not merely receive a winner label.

Synthesis introduces new decisions

An integrator may combine Candidate A's API with Candidate B's validation strategy. The combined result can create conflicts neither candidate had. Imports move, error behavior changes, or duplicated assumptions collide.

That is why integrators produce complete finals on separate neutral branches. Each final must pass the gate again. "Both ingredients passed" is not evidence that the mixture passes.

Pass two compares finals with the baseline

The second jury receives independently randomized neutral aliases. It evaluates every integrated result against three references:

  1. the frozen contract;
  2. the original baseline and regression guards;
  3. the pass-one findings the integration was meant to address.

This prevents a polished synthesis from winning simply because it looks more comprehensive. The final must preserve what already worked and improve the requested slice.

Before and after stay linked

The pass-2 packet carries what the first pass learned: which criteria separated the candidates, which grafts the judges ordered, and where they disagreed. It also includes the candidate baseline, integrator dispositions, re-run gate evidence, and the final diff. A judge can trace a strong error path from its candidate origin into the final, or notice that it vanished.

The rubric does not move

Integration is not permission to rewrite the Definition of Good. If synthesis reveals a defective criterion, the run routes back through an explicit contract revision.

Fresh QA still follows selection

Jury pass two chooses the strongest final. It does not make the final ready to ship by itself.

A fresh QA review looks for obvious blockers, followed by the required deeper QA audit. These sessions inspect the selected final without carrying the tournament's momentum. A late issue routes back to bounded repair, not around the gate.

Two passes make the tournament accountable at both creative moments: invention and synthesis. Fresh QA makes sure the final result, not the competition story, earns the handoff.

Source ledger

  • Fusion planning model, section 6.
  • Fusion MVP PRD, FR-12.
  • Integrator tournament summary, section 14.