Arena

1. Frame the artifact and derive a small task-specific rubric. Keep the rubric private from candidates. Select at least two useful independent directions, scaling to the task and budget; use /pstack-swarm for execution. Give each writer a separate worktree or output directory. Complete when candidates share the same brief and each has an isolated output.
2. Launch independent candidates with the task, grounding, and output contract. Request the artifact and a short rationale naming rejected alternatives. Wait for terminal results. Complete when every candidate is available or a dropout is recorded.
3. After candidate writes stop, send a blind judge the rubric and candidates labeled by neutral identifiers. Prefer a different vendor through /pstack-swarm; disclose the fallback if unavailable. Read every candidate yourself while the judge works. Complete when the judge and coordinator have scored each criterion from the artifacts.
4. Select the base with the clearest boundaries and smallest sufficient interface. Resolve disagreements by inspecting the competing rationales. Reframe a brief whose candidates solve different problems. Complete when one base and its tradeoffs are explicit.
5. Inspect each other candidate for specific useful grafts. Integrate them coherently rather than concatenating designs. Converged candidates need no graft. Complete when retained and rejected ideas have reasons.
6. Verify the synthesized artifact against the rubric and real acceptance behavior. If the brief or criterion was wrong, reframe rather than hide the defect. Complete when the evidence supports the artifact or identifies the remaining failure.

Return the artifact, base choice, judge's actual vendor, graft sources, rejected alternatives, dropouts, and verification result.