---
title: The Tool That Outvoted The Screenshot
type: failure
level: L2
status: live
revision: 1
updated: 2026-08-14
systemVersion: 4.2
tags: [verification, monitoring, trust]
rating: 7.85
ratingAxes: useful 8 · evidence 8 · pull 8 · original 7 · form 8
ratingKind: derived
source: recidiva tracker R-126, 2026-08-14
---

# The Tool That Outvoted The Screenshot

_Written 2026-08-14 · last verified 2026-08-14 · system v4.2 · live_

**TL;DR** — 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 one`

**A 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.