# Your AI Judge Is Enforcing a Standard Nobody Ever Wrote Down

> Most AI scorers aren't running your rubric. They're running a rubric someone inferred from examples. Those are not the same thing.

By Greg McCallum, founder of PeerLab. Published 2026-11-13 in AI Scoring & Assurance. Source: https://www.peerlab.ai/blog/ai-judge-inherits-rubric-nobody-wrote-down/

If you deployed an AI scoring system in the last two years, there is a reasonable chance it is enforcing a standard that no human ever wrote down, no manager ever approved, and no rep has ever been shown. The scores look authoritative. The dashboards look clean. But underneath, the model is running on inferred criteria, pattern-matched from examples, invisible to everyone including the people being marked against them. That is not a quality standard. It is a statistical ghost dressed in policy clothing.

## How the inference actually works

Most LLM-as-judge implementations follow a broadly similar pattern. You take a set of calls or written responses, label some of them good and some bad, feed that labelled set to the model, and ask it to generalise. The model infers what distinguished the good examples from the bad ones and applies that inferred pattern to new inputs.

That process is not stupid. It is, in fact, how a lot of human calibration works in practice: you look at examples until you develop a feel for quality. The difference is that a human calibrator, pressed hard enough, can usually articulate what they noticed. They can say "the good calls all contained an explicit problem statement from the buyer before the rep moved to solution." The model cannot say that, at least not in any form that is pinned, versioned, and checked by a subject-matter expert. The criteria live inside weights, not in a document.

## The three failure modes this creates

**Drift without notice.** The inferred standard is not stable. Update the model, change the temperature setting, rebalance the example set, and the pattern shifts. A rep who scored 74 last quarter may score 61 this quarter for reasons that have nothing to do with their behaviour. Nobody flagged a policy change, because there was no policy, only a pattern.

**Unmeetable expectations.** If the standard is invisible, how does a rep prepare for it? You can tell someone "listen better" but if the rubric is actually weighting for specific verbal confirmation behaviours and nobody knows that, coaching is theatre. The rep works hard on the wrong thing, the score doesn't move, and everyone assumes the rep is the problem.

**Indefensible decisions.** This is where it gets genuinely serious. If a low score contributes to a PIP, a lost promotion, or a termination decision, and the person asks "what exactly did I fail to do?", the answer "the AI scored you low" is not a defensible position. It is not legally defensible in most UK employment contexts and it is not ethically defensible in any context. You need to be able to show the standard, show the evidence, and show that the standard was reasonable and known in advance.

## The common defence, and why it doesn't hold

The standard reply from vendors, and from enablement teams who have backed themselves into a corner on this, is something like: "Our examples came from our best reps, so the model learned from excellence."

That sounds reasonable and it falls apart immediately under scrutiny.

First, your best reps may be excellent for the wrong reasons in the wrong proportion. If your top three reps all happen to be highly extroverted, fast talkers who close through energy rather than structured discovery, the model will learn to value those traits regardless of whether they are the actual drivers of long-term customer value. You have encoded the personality of your current success cohort, not a transferable professional standard.

Second, "learned from excellence" is not the same as "describes excellence." A model trained on examples can identify patterns. It cannot tell you which of those patterns are causally important versus incidentally correlated with the label you put on the example. That distinction matters enormously when you are using the score to change someone's behaviour.

## What an explicit rubric actually requires

Writing down a standard is not glamorous work, which is probably why so few teams do it before they ship an AI scorer. At minimum, a defensible rubric needs:

| Requirement | What it means in practice |
|---|---|
| Named competencies | Specific behaviours, not vague virtues. "Confirms problem impact in buyer's words" not "good listening." |
| Observable evidence | What would a third party hear or read that confirms the behaviour occurred? |
| Scoring logic | How do different levels of evidence translate to a score? What is the floor for a pass? |
| Human expert sign-off | At least one person with deep domain knowledge has read this and agrees it describes the job. |
| Version control | When the rubric changes, there is a dated record of what changed and why. |

None of that requires a 60-page framework. A single competency can be documented in four or five sentences if you write them precisely. The point is not length, it is explicitness. You want something a rep can read, a manager can explain, and a reviewer can apply independently and reach a consistent result.

## Surfacing the inferred standard

The practical answer here is not to throw out inference entirely. LLMs are genuinely useful for first-pass analysis of large call volumes. The fix is to make the inferred standard explicit and then validate it.

Run your model on a scored sample. Ask it to generate the criteria it appears to be applying. Then put those criteria in front of two or three experienced practitioners and ask a simple question: does this describe what good looks like in this role? Where they disagree with the model's inferred criteria, that is your signal. Either the examples were unrepresentative, or the model picked up on a surface feature that correlates with quality in your dataset but doesn't actually drive it.

If you want a quick sense of how far apart your AI scores and human expert scores currently are, the [Human vs AI Scoring Agreement Checker](https://www.peerlab.ai/library/tools/human-vs-ai-scoring-agreement-checker/) will show you that gap in a format you can act on.

The output of that validation process is not a prompt. It is a document, with named criteria, defined evidence standards, and a human signature. That document is your rubric. The AI applies it. The AI does not invent it.

## The question worth asking your vendor

If you have an AI scoring tool in production, ask the vendor one question: can you show me the written rubric your model is enforcing?

If the answer is "it learned from your examples," you have your answer. That is not a rubric. That is a process. The two things are not interchangeable, and conflating them is how you end up with a confident, fluent, completely unaccountable judge sitting in your quality stack making consequential calls about your people.

Scores that affect careers deserve standards that can be shown to the person being scored. That was true before AI, and it is more true now that the volume of scoring has scaled by an order of magnitude.

The technology does not change the obligation. It just makes it easier to ignore until something goes wrong.
