Skip to content

Race

Race runs N attempts at the same change, scores each on a fixed rubric, and verifies the top attempt before anything is committed. Losing attempts are never committed. It trades Forge’s cross-review for raw speed plus a hard verification gate — the move for high-stakes single changes where you want one correct, proven result.

ForgeRace
Middle stepRound-robin cross-vendor reviewNone — straight to scoring
DecisionJudge ranks using peer reviewsFixed rubric + hard verification gate
What commitsWinner, after you see the diffsTop attempt only if it passes verification
Best forBest implementation of a specA correct, proven result, fast

By default Race uses distinct vendors — the best shot from each family. You can instead run the same vendor across several independent attempts (a temperature/seed spread) if you want N tries from one model. Say which you prefer; the default is distinct vendors.

The rubric is defined and shown to you before any attempt runs — it must be objective. A typical rubric: meets the acceptance criteria, tests pass, no type/lint errors, handles the named edge cases, and stays simple (no gratuitous churn).

Each attempt runs in its own isolated git worktree (the same isolation Forge uses), implementing the task against the rubric. No attempt sees the others.

Every attempt is scored against the rubric and ranked. The top-ranked attempt is then verified — the test/build/lint flow the rubric named is run against it in its worktree:

  • Passes → it’s committed to your active branch (after you see the diff and scorecard).
  • Fails → it’s dropped and the next-ranked attempt is verified, and so on down the ranking.
  • All fail → nothing is committed; you get the ranked attempts and their failures to decide.
  • Verification is the contract — nothing lands unless it passes the rubric’s checks.
  • Never commits to main — verified winners land on your active branch only, after you’ve seen the diff.
  • Worktrees are cleaned up at the end; losing attempts are never committed.

Natural language triggers:

  • “Race the vendors on this”
  • “Best verified attempt at X”
  • “Have the models compete and commit the one that passes”

Explicit harness: select Tribunal Conductor from the harness picker, then describe the change and (optionally) the rubric.

  • Needs runnable verification — Race’s value is “proven, not just plausible”. If there’s no test/build/lint that can verify the change, prefer Forge’s peer review instead.
  • The rubric must be fixed up front — a rubric written after seeing the code isn’t a fair contest.
  • Cost — N attempts plus verification. Ptah announces the attempts and cost before spending.