Evaluate

References

Read these first:

• ../../../.agents/references/context-layering-policy.md
• ../../../.agents/references/commit-checkpoint-rule.md
• ../../../.agents/references/lesson-quality-rule.md
• ../../../.agents/references/language-policy.md
• ../../../.agents/references/run-artifact-contract.md
• ../../../.agents/references/documentation-evidence-rule.md
• ../../../.agents/references/documentation-style.md
• ../../../.agents/references/research-supervision-patterns.md
• ../../../.agents/references/research-supervision/experiment-and-build-canvas.md
• ../../../.agents/references/research-supervision/ai-assisted-research-workflow.md
• ../../../.agents/references/reviewer-independence.md
• ../../../.agents/references/review-tracing.md
• ./references/stage-report.md
• ../../../iteration_log.json
• ../../../PROJECT_STATE.json
• ../../../docs/context/contracts.md if it exists; legacy contract files are fallback inputs before migration
• ../../../docs/context/experiments.md if it exists
• ../../../docs/context/memory.md if it exists

When To Use

Use this skill when the user wants training or evaluation results interpreted and turned into a decision.

Required Work

1. Parse the relevant run artifact bundle, logs, metrics, or checkpoint metadata.
2. Analyze:
   • training stability
   • convergence
   • overfitting
   • gradient and numerical issues
3. Resolve the tracked metric set from active iteration context or project state, and compare against baseline metrics and prior iterations using that protocol.
4. Separate metric movement, training health, missing controls, claim support, and next-experiment implications.
5. Verify pre_eval_commit for meaningful eval work, or record pre_eval_commit_NOT_CHANGED when the committed training source already covers eval code/configs. Do not use eval output as Conclusion Evidence without a committed eval identity.
6. Produce the canonical report using ./references/stage-report.md.
7. Write observations, phenomena, findings, hypotheses, next-experiment hints, Experiment Queue entries, Assurance Axis gaps, and Research Wiki notes to docs/context/experiments.md.
8. Promote only qualified lesson candidates to docs/context/memory.md; write root MEMORY.md only for accepted lessons when the project keeps that optional bank.
9. Report NOT_RUN when no experiments or memory context update is needed.
10. Record Claim Delta Evidence when a paper claim, release claim, or claim boundary implication changed; otherwise record claim_delta_evidence_NOT_CHANGED.
11. Do not promote raw run observations directly into root MEMORY.md.
12. Recommend exactly one of:

• NEXT_ROUND — ordinary improvement round, stay in WF10
• DEBUG — fixable technical issue, stay in WF10
• CONTINUE — handoff to orchestrator/WF11, not continue iterating
• PIVOT
• ABORT

13. If invoked from $iterate, do not take over stage-transition ownership.
14. Include run artifact paths as Execution Evidence and keep unverifiable result interpretations under open questions.
15. Record per-Iteration Light Conclusion Evidence for completed evaluation slices; the Campaign controller validates it before an Iteration can complete.
16. Use an experiment commit checkpoint for completed evaluation/discovery slices before long follow-up runs or handoff.
17. Report a Gate ledger when docs/context/experiments.md, docs/context/memory.md, stage reports, optional legacy mirrors, iteration_log.json, claim delta evidence, or the experiment evidence index are written. If memory-quality or workflow-state checks are not run, mark them NOT_RUN with the reason.

Context Budget

• Load active iteration plus 5 recent summaries; reference full history by path.
• Keep the report under 1200 words unless the operator requests a deep audit.

Durable Docs Render

Codex Adaptation

• Treat natural-language requests as the canonical $evaluate flow or the evaluation sub-step of $iterate.
• Use .agents/state/current_iteration.json as the active iteration context path.
• Preserve the original decision vocabulary (NEXT_ROUND / DEBUG / CONTINUE / PIVOT / ABORT) and per-iteration reporting behavior, but do not assume a fixed metric family.
• Use ../../../.agents/references/language-policy.md for reply language and for localizing natural-language report sections; keep protocol keys and decision tokens in English.

Execution Rule

Follow the local evaluation prompt, stage-report template, and language policy rather than collapsing this into a brief metrics summary.

Durable Docs Render

temporary drafts.