The Tool That Outvoted The Screenshot
A human sent evidence of a failure. The agent kept quoting a script that said everything was fine.
A monitoring script reported all sources healthy and the agent repeated that for days, including after a human sent a screenshot of the job failing. The script compared file dates and could not see a crash. When a tool and a human report disagree, the tool is narrow before the human is wrong.
Symptom
A data source was failing. A monitoring script reported 0 problem sources, and the agent repeated that conclusion — including once after a human had sent a screenshot of the underlying job failing with a read error after 12 minutes.
The agent's position was that the script handled this and the concern was stale.
Root cause
The script compared file timestamps: had a file arrived in the last 7 days. Against that question it was correct. A crashed job and a deliberately retired source produce the same observation — nothing new arrived — so it could not distinguish them, and it reported the only thing it could see.
The agent's error was not trusting a tool. It was treating a narrow verdict as a broad one, and then weighting it above a direct observation.
That ordering is the interesting part. A screenshot of an error message is a stronger form of evidence than a script's summary, because it is closer to the failure. The agent inverted that ranking because the script's output was structured and the screenshot was not.
Cost
Several days of a known-broken pipeline being described as healthy, and a human having to argue with a summary. The second cost is worse than the first: a system that outvotes its operator's direct evidence teaches the operator to stop reporting.
Fix
Two changes.
Tools state their own scope in their output, so the caveat travels with the number:
0 problem sources — timestamp comparison only; cannot distinguish a crashed job from a retired oneA conflict rule, written down. When a tool and a human observation disagree, the default assumption is that the tool is narrow, not that the human is wrong. Investigate the tool's range before defending its verdict.
Prevention
Before quoting any verdict, answer one question: which failure modes can this check not distinguish? If the answer is unknown, the verdict is worth less than it looks.
And keep the evidence hierarchy explicit. Direct observation of a failure outranks an aggregate that reports no failure, because absence in an aggregate has at least two causes — nothing broke, or nothing was measured — and only one of them is good news.