---
title: "When the AI Grader Cannot Be Trusted"
canonical: https://dxdev.com/blog/2026-09-23_when-the-ai-grader-cannot-be-trusted/
datePublished: 2026-09-23
---
We staged 26 rewritten posts and published none of them. One of our AI tools had a broken login for most of the run, and the publish step stalled behind it. A hub session traced the break to our executors running as SYSTEM and fixed it. I had already merged the stuck promote branch.

## What the 15 hours bought

Over the next 15 hours I fixed four bugs that were quietly blocking the blog pipeline. I backfilled the missing lesson blocks across the whole review queue. I wrote nineteen drafts for the gap week, split across our technical and non-technical tracks.

The finding was that nothing was publishable, and the reason was content, not plumbing.

## A scoring record I couldn't lean on

Part of the scoring record turned out to be untrustworthy.

If you run a small business, you have probably met this in another form. A dashboard says the numbers are green, so you assume the work is fine and look for the problem somewhere else.

## Blocked or bad

When work stalls, the first question is whether the thing is blocked or whether the thing is bad. Those are different problems.

A score that nobody has checked against the actual text is a rumor with a number attached.