Some failures don't throw an error — they just quietly don't happen. A route file added without updating the route registry. A toggle flipped in the database that a cached frontend never reads. A deploy that ran but published nothing because a build step was silently skipped. An agent that reads untrusted content — a scraped page, a user upload — and quietly follows an instruction buried inside it, no crash, just the wrong action taken. These all pass every naive check because nothing crashed.
The catch is asking a different question than "did it error." Ask "what had to happen for this to actually take effect, and did I check that specific thing" — not the symptom, the mechanism.
Run the seed: given a "shipped, no errors" report, list two ways it could still be silently broken, and name the specific check for each.