Evolution Decide

What to Do

1. Read All Artifacts — Open implementation.md and benchmark.md to review what was implemented and how it performed. Open iteration.json for the full context including hypothesis, state, and metrics.

2. Evaluate Against Acceptance Policy — Determine the outcome based on:

   • Accepted: The implementation succeeded, the benchmark evidence satisfies the documented policy, benchmark.sufficientForPromotion is true with a statistically meaningful improvement over the baseline, and the candidate can be promoted under the versioning policy.
   • Rejected: The implementation succeeded and benchmark evidence is strong enough to show the candidate is weaker than or equivalent to the baseline. The candidate is discarded.
   • Inconclusive: The implementation succeeded but the benchmark evidence is insufficient for a clear accept/reject decision (for example: only screening evidence, weak signal, high variance, or confirmation still required). The candidate may be refined in a future iteration.
   • Failed: The implementation, correctness gate, or benchmark infrastructure failed (build error, test failure, benchmark crash). No evaluation of the candidate's merit is possible.

3. Write decision.md — Record the decision with:

   • The final outcome (accepted, rejected, inconclusive, failed)
   • The reasoning for the decision
   • Key evidence cited from the benchmark results, including policyStage, completed games, and whether benchmark.sufficientForPromotion was satisfied
   • For an accepted outcome, the promotion metadata required by the versioning policy (previousVersion, promotedVersion, baseline refs, and version-artifact path)
   • Any recommendations for future iterations

4. Update iteration.json — Follow the canonical state-machine contract in tasks/prd-wiggum-evolution-loop.md:

   • Enter the decision phase only from benchmarked by transitioning to deciding before choosing an outcome
   • Set the final state to exactly one of "accepted", "rejected", "inconclusive", or "failed"
   • Keep stateMachine.currentPhase aligned with the final state
   • Add decision object with:
     • outcome — one of the allowed final states
     • reasoning — explanation of the decision
     • evidence — key metrics that supported the decision, including benchmark policy fields used for the call

Decision Format

The decision artifact should follow this structure in decision.md:

# Iteration N Decision

## Outcome

One of: accepted, rejected, inconclusive, failed.

## Reasoning

Explanation of why this outcome was selected based on the evidence.

## Evidence

- Implementation: summary of what was changed
- Benchmark: key results (games completed, win rate, ELO estimate)
- Policy: which acceptance criteria were met or not met

## Recommendations

Suggestions for future iterations (e.g., "retry with more games", "explore related idea from iteration X", "abandon this direction").

Allowed Final States

The final-state rules come from the iteration state machine in tasks/prd-wiggum-evolution-loop.md. The decision phase must not skip directly from implemented to a final outcome.

Scope Constraints

• Only write decision.md and update iteration.json; do not modify engine code, benchmark results, or other iteration artifacts
• Do not persist or discard candidate changes in the codebase — that is handled by the orchestration script
• Do not bump version numbers — that is handled by the orchestration script on accepted candidates

Output Contract

After this skill runs, the following must be true:

• decision.md exists in the iteration directory with the outcome and reasoning
• iteration.json has been updated with one of the four allowed final states: accepted, rejected, inconclusive, or failed
• iteration.json includes a decision object with outcome and reasoning fields

References

• Worker guidance: .claude/evolution/CLAUDE.md
• Evolution loop: scripts/evolution-loop.sh
• Benchmarks: scripts/benchmark-version.sh
• Benchmark policy: see US-011 benchmark policy contract
• Versioning policy: see US-013 versioning policy contract in tasks/prd-wiggum-evolution-loop.md
• Iteration state machine: tasks/prd-wiggum-evolution-loop.md