AI Version Regression Explainer: What Regressed, in Business Terms

'POST /rulesets/:id/regression' replays real recorded decisions through both versions and reports: changed outputs (which keys moved, in how many scenarios, first example), rule churn (rules that stopped or started firing), equivalence verdict, and coincidental agreement.

**Coincidental agreement** — same output, different rules — is counted separately from an identical run. A migration that claims parity on outputs but produces them through different rule paths is not a behavioural equivalent; the next out-of-distribution input might diverge. This distinction is what the replay measures and what a structural diff cannot.

**The AI step** is called after the measurement: 'POST /ai/review-diff' is given the structural diff and observed fire counts. The answer is joined onto the structural diff — a changed rule the review ignores is a visible gap; a rule the review names that's not in the diff is an orphan.

**CLI integration**: 'dm regression run --ruleset loan-eligibility --base v11 --target v12 --corpus last-30-days' 'dm regression assert --run <runId> --max-changed-scenarios 0'

Exits 0 if assertion holds, 1 if it fails. Run URL printed to stdout for PR annotation.