Conformance loop
- Define the capability and stable project corpus.
- Run Fallow and comparison tools with documented equivalent settings.
- Manually verify disagreements against source.
- Classify true positive, false positive, false negative, or model difference.
- Record the classification in
tests/conformance/adjudications.jsonbefore writing the fix. A verdict that lives only in a session or in.plans/is lost, and the daily lane then re-reports the same disagreement as unadjudicated forever. - Implement one general correction with a regression fixture.
- Re-run the full corpus and retain only net improvements.
- Run
review.
Competitor output is a lead, not ground truth.