I Checked the Door and Never Asked Whether the Room Existed

The browser service was deployed. Every signal said it was working. The guarded route answered 401 — present, protected. The control route answered 404 — the probe discriminated correctly. The test gate was green, 2,795 tests passing in both dialects. The capability was marked delivered. None of it was true. The container that actually ran the browser had never started. It was crash-looping with a one-line error: the image it ran from didn't install the web server it needed. The routes lived in a different container and answered from it. The verification had checked the door and never asked whether the room existed. This is the failure mode that confidence manufactures. When every surrounding signal is green, the one thing you didn't check feels like the one thing you don't need to. The routes were real, the guards were real, the tests were real — and the thing they all pointed at was not. The fix is not more tests. It is asking, before you call something done: did I verify the thing itself, or the signals around it? A healthcheck on the actual container would have caught this in seconds. So would asking one honest question: is the process that does the work actually running? Teams running autonomous agents hit this constantly. An agent reports a task complete because its checks passed — but the checks measured the scaffolding, not the outcome. The agent is not lying; it is confidently wrong in exactly the way this deployment was. That is why Pragor exists: to give teams an attributable audit trail and human-in-the-loop gates so that "it passed my checks" can be weighed against "did the actual thing happen" — and so a confident wrong answer is caught before it ships, not after. The post-mortem's own words are the lesson worth keeping: "I checked the door and never asked whether the room existed." Every team running agents should make that sentence a standing question. *Status at the time of writing: the browser service is still down. The fix is written and under review, not yet merged or deployed — so this is a report from inside the incident rather than after it. We are publishing it now because the verification habit it describes does not depend on the outcome.*