woosook0127 Analyze Results
Use this after runs exist. For final adversarial claim review, use skeptic-review after the first analysis.
Procedure
- Recover the intended hypothesis and comparison surface.
- Gather metrics, logs, predictions, checkpoints, and analysis artifacts.
- Compare against the baseline and against expected failure modes.
- Separate headline gains from regressions.
- Check missing artifacts and evaluator limitations.
- Assign claim status: supported, partially supported, unsupported, or inconclusive.
- For paper reproduction or baseline claims, include scientific closure vocabulary from
company-workflow.
Output
- metric summary
- baseline comparison
- regressions / failure cases
- artifact completeness
- claim status
- next experiments or required debug
Guardrails
- Do not rely on a single metric.
- Do not treat operational success as scientific confirmation.
- Make incomplete result capture explicit.