mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-10-05 22:35:57 +02:00
Document the methodology for validating behavior-changing skill prose edits: paired old-vs-new blind injection to separate demonstrated improvement from no-regression, reading non-discriminating results honestly, standardizing field names for cross-skill gates, and proving a parity test fails on one-sided drift. Add reciprocal Related links from the three adjacent skill-eval docs.