We seed small bugs into a file and see whether its tests notice. On platform/audio/narrate.py, the engine that voices our podcasts, the tests caught 3 of 8 (37.5%).
The five that escaped were real holes: the one-core fallback when the CPU count is unknown, the rule that the English-only phoneme converter must hand other languages to espeak, the cache key for a missing model file, the one-unit-at-a-time pool scheduling, and ffmpeg failures being silently ignored instead of raising. Each now has a test that fails with the bug and passes without it, and a re-run catches all eight.
The lesson is the usual one: tests that exercise the happy path say little about the guards around it.