How to reason about it
- Freeze the corpus inventory before comparison and classify every item as agree, mismatch, or not-comparable.
- Review mismatch inputs as behavior changes; do not regenerate old receipts to erase history.
- Preserve runner identity, rule bytes, input bytes, and result observations for reproducibility.
Compare with a baseline
SH
lispex vouch replay gallery \
--against <baseline-version-or-receipts-dir>Read every reported mismatch and not-comparable case. A clean summary is useful only when the corpus and baseline were chosen before the outcome was known.
A common mistake
Not-comparable is a disclosed boundary, not a mismatch silently counted as agreement.
Keep going
Keep the Vouch overview nearby: it separates authentication, request binding, current execution agreement, and a local grant.