see repo health, trust the score, fix with a PR. An analyzer that fails never renders as healthy.
Every score carries its evidence and its confidence, an analyzer that fails reports degraded instead of clean, and a fix only ever arrives as a pull request.
A single composite number with failed analyzers counted as zero findings. Rejected because could-not-look and nothing-wrong would render identically, and the product would lie by default.

Measured: the scan's own scores, with the composite cap and the confidence shown. Claimed: that a fix improves the score. The product re-scans the fix branch and posts the realized change, even when it misses the projection.