Picking the best candidate leaves value on the table. Combining candidates creates new failure modes: synthesis can drop grafts, flatten dissent, or introduce regressions. Fusion's answer is to make integration itself a tested part of the tournament. Configured integrators produce competing complete finals, each final is re-gated, and both return to the jury for pass 2.

Candidate competition explores implementation space. Integrator competition explores synthesis space. The best individual solution and the best combined final are different optimization problems, and Fusion keeps both.

An integrator receives the same frozen contract, anonymized candidate artifacts, gate evidence, jury rationale, and ordered graft guidance. It must produce one coherent implementation, not a patchwork report.

Synthesis creates a new artifact

The final can differ materially from every candidate. A useful feature from runner-up B may enter candidate A's structure. A conservative error path from A may replace B's looser one. Pass-one judges may prefer candidate A overall while finding that candidate B has safer error handling and clearer tests; picking A unchanged would discard useful evidence, and blindly merging B into A could introduce new defects.

Those choices need provenance. Each ordered graft should be preserved, modified with reason, or explicitly rejected. Silence is not a disposition.

Competition tests the integration strategy

Two integrators may interpret the same evidence differently. One may prefer a minimal graft. Another may find that a shared abstraction resolves both candidates' weaknesses. Synthesis involves taste and risk: both approaches can be defensible, and each can fail differently.

Running more than one integrator prevents the first synthesis plan from becoming the unchallenged final. The integrators work from shared neutral inputs, not private conversations with favored candidates. Both finals must stand alone, pass the objective gate, and face the same second jury. The selected result is the final artifact, not the reputation of its integrator.

Every integrated result starts trust from zero

The final is a new commit with its own SHA. Fusion reruns the protected gate and binds new evidence to that artifact. Candidate evidence remains provenance, not inherited proof.

The second blind jury then compares the integrated finals. It checks whether each one:

  • satisfies every criterion;
  • resolves pass-one findings;
  • preserves baseline behavior;
  • stays within scope;
  • avoids adding unnecessary risk.

Once pass 2 and fresh QA select a final, the tournament stops. The result becomes ready to ship, and a human reviews the evidence and authorizes the consequential remote action.

Integrator competition is not extra ceremony. Combining good ideas is itself hard engineering, so the combination deserves alternatives and proof.

Source ledger

  • Integrator tournament summary, sections 2, 4, 5, and 11.
  • Fusion MVP PRD, FR-18 and section 9a.