“It came back with a rating — it went through the evaluator, and the evaluator rated it 4.9 out of 5. And I’m looking at the quality of the writing, I’m going, this is not 4.9 out of 5. But against the rubric, or against the evaluation criteria that that evaluator agent developed over time, it was 4.9 out of 5.” — Lou
Session context: 2026-07-23_Mastermind — Lou opened the session with a post-mortem on a batch of long-form articles he’d produced and disliked. The evaluator disagreed with him, confidently, and it was right about its own criteria.
Core Idea
A trained rubric is the standard answer to AI slop, and it works — until the rubric becomes the thing that needs auditing. Lou’s evaluator agent had accumulated its evaluation criteria over many runs. Every one of those accumulations was reasonable at the time. The compound result was a scoring instrument that certified writing its owner found unreadable, at 4.9 out of 5, with no hesitation.
This is not the evaluator malfunctioning. It is the evaluator working exactly as built and measuring something that quietly stopped being what you cared about. The failure has no error message. A broken pipeline throws; a drifted rubric ships — and it ships with a number attached, which is worse, because the number suppresses the doubt you would otherwise have had.
Lou named the distinction that makes this legible, and it became the title of the article he published about the episode: did you do it right, or did you do the right thing? A rubric can only ever answer the first question. It measures conformance to criteria. Whether those criteria still describe good work is a judgment that lives outside the loop, and the more you automate, the fewer occasions you have to make it.
The compounding risk is that this sits underneath everything else you’ve automated. A drifted evaluator doesn’t just pass one bad artifact — it becomes the fitness function for a self-improving system, and the system will optimize toward it faithfully. You end up with a pipeline that gets measurably better at producing work you like less.
The tell is the gap itself: your gut says one thing, the score says another. Most operators resolve that gap in favour of the score, because the score looks like evidence. The correct move is the opposite. When your reaction and your rubric disagree, the rubric is the suspect. You are the ground truth; the rubric is a lossy compression of you, taken at an earlier date.
Practical Application
Run a calibration pass on any evaluator you use regularly. Pull five artifacts your evaluator scored above 90% and read them cold, without looking at the scores. Rank them yourself, worst to best. Then compare your ranking to the machine’s. Where the orderings disagree, open the rubric and find the criterion responsible — you will usually find one line that made sense the day it was added and has been quietly over-weighted ever since. Do this every time the evaluator has learned something new, not on a calendar.
Related Insights
- Insight - The Output Quality Gate — Train a Rubric, Not a Prompt — established that training a rubric beats tuning a prompt; this is that pattern’s failure mode.
- Insight - The Quality Gate Pattern — Embed 9-10 Self-Evaluation at Every Pipeline Handoff — self-evaluation at every handoff is exactly what propagates a drifted rubric through a whole pipeline.
- Insight - The Self-Improving Skill Loop — Have the Skill Learn From Every Use — the mechanism by which the criteria accumulated in the first place.
- Insight - Your Growing Trust in AI Is the New Attack Surface — a confident score is a trust signal, and this is what it costs when it’s wrong.
- Insight - Your Own Rules Are the Bug — Use One Agent to Debug Another — the same session’s companion failure: instructions you wrote and forgot, faithfully executed.
Evolution Across Sessions
Builds directly on Insight - The Output Quality Gate — Train a Rubric, Not a Prompt (2026-06-04), which established that a trained rubric outperforms prompt-tweaking for consistent output quality. The new development is the second-order problem: a rubric that learns is a rubric that drifts, and drift is invisible because the output is a confident number rather than an error. This adds a required maintenance step to the quality-gate pattern — periodically re-anchor the evaluator against your own cold read — and it establishes a baseline for a theme this room will keep hitting as more of the pipeline self-improves: the measurement instrument needs its own audit.