A tournament that ends with "here are three candidates, pick one" has exported its hardest job to you. Fusion's answer is a distinct role with a narrow mandate: the integrator.
The mandate
The integrator starts from the jury's pass-one verdict: a winner, plus specific ideas from the losing candidates that judges flagged as worth keeping. In a neutral worktree, it produces one complete final:
- the winning candidate as the base;
- approved grafts folded in, and only approved grafts;
- the full objective gate re-run on the combined result;
- provenance recorded for every criterion: which candidate contributed what.
Then the final goes back to the jury for a second blind pass. The integrator does not grade its own work, does not push, does not merge, does not ship. It builds and hands off.
Why grafting needs a referee
"Combine the best of both" sounds obvious and fails constantly when done casually. Two candidates that each pass the gate can produce a merged result that passes nothing; code is not lego. Three rules keep it honest:
- Grafts must be approved: a judge named the idea worth keeping. The integrator has no license to freelance.
- The gate runs again on the final. A graft that breaks the combination is discovered by machinery, not by your reviewer at midnight.
- The second jury pass grades the final as a stranger's work, blind like everything else.
What you get out of it
One branch, one diff, one receipt, with a provenance trail that says where each piece came from and proof that the whole passed together. The tournament happens so that you never have to think about the tournament.