The regression runner's numbers wobbled run-to-run — not because the system changed, but because one step was non-deterministic. A gate that fires on its own noise gets muted.
Passing tests suddenly failed right after a code refactor. The refactor was innocent — a gitignored database had silently gone out of date.
The system sometimes matches on shared words rather than shared cause — with high confidence. A known limitation kept visible on purpose, because it's the exact failure the project exists to guard against.
Building the first measured baseline for the diagnostic agent caught a real bug 30 unit tests had missed: a safeguard that let fabricated citations through as long as one real one rode alongside.
Every test passed and every test was correct — and the suite took 15m 46s, which in practice means nobody runs it. Getting it to 8.5s.