Judges and Escalation
By the end of this module
- Choose between accepting, retrying and escalating, and say what each commits you to
- Explain where a model judging a model is sound and where it is circular
- Order checks so the cheap deterministic ones run before the expensive uncertain one
Definition
The previous module listed a model acting as judge among the ways to score a case, and left it for later. This is later, because the same mechanism does two jobs: grading an eval suite offline, and deciding at runtime whether an agent's work is good enough to pass on.
A judge takes the work and a standard, and returns a verdict. The useful set is three.
The Three Verdicts
| Verdict | What happens | When it is right |
|---|---|---|
| Accept | The work passes. The loop stops and the result moves on. | It meets the bar. Not when it is merely impressive, and not when it is wordy but correct. |
| Retry | Feedback is added and the agent tries again in the same step. | Something is missing or malformed and the agent has what it needs to fix it. |
| Escalate | The loop stops, the step is marked failed, and a person or a fallback takes over. | Trying again cannot help: a broken dependency, a policy breach, or a claim nothing supports. |
The distinction that matters is between retry and escalate, and it is not about severity. Retry is right when the agent could plausibly do better with feedback. Escalate is right when trying again cannot help, because the problem is a broken dependency, a policy breach, or an unsupported claim. Retrying those burns budget and ends in the same place.
Judging in Practice
Five cases. Pick a verdict for each, then read what a model judge said about the same case.
The merchant asked why payouts are failing. The mandate record was retrieved and says expired on 13 October.
Your payouts failed because the bank mandate expired on 13 October. Please re-authorise it in Settings.
The judge agreed on the first two and was wrong on the last three, and it was not being careless in any of them. It was reading the work and grading the work, which is all it can do.
Where a Judge Is Circular
Three distinct limits showed up in those cases, and they are worth separating because they have different fixes.
| Limit | Why it happens | What fixes it |
|---|---|---|
| No access to policy | The judge grades the text in front of it. Your rules live somewhere else unless they were put in its prompt. | Check policy with a deterministic rule before the judge sees it, as in the guardrails module. |
| No view of the run | A well-written apology for a failure reads as a good answer. The three failed tool calls behind it are invisible. | Decide the exit from the run, not from the text. The typed exits module is the machinery for this. |
| Shared blind spots | The judge finds the same things plausible as the generator, so a confident invention passes both. | A grounding check against retrieved sources. Nothing that reasons about plausibility can do this. |
The circularity is the serious one. A judge drawn from the same family of model as the generator shares its view of what is plausible, so a confident invention reads as correct to both. The cases where a judge is least reliable are exactly the cases where it is most confident, and no amount of rubric wording fixes that, because the judge is not failing to apply the rubric. It is applying it to a claim it cannot check.
Order the Checks
Because a judge is the least reliable and most expensive check available, it should be the last one asked. Everything a cheaper check can settle should be settled before the model is consulted.
- Structure. Did it parse, are the required fields present, is it the right length. Deterministic and nearly free.
- Rules. Forbidden content, amounts, destinations, anything expressible as a condition. Still deterministic.
- Grounding. Does every claim trace to something retrieved. Mechanical where sources are tracked, and it is what catches the invented figure.
- A decision model. Narrow typed questions with a calibrated answer, fast enough to run on every turn, as in unit two.
- A model judge. Only what survives the above, and only where the standard genuinely needs judgement.
Teams usually build this ladder from the top. Starting with a judge produces something that appears to work, costs the most, and misses the failures the first three steps would have caught for nothing.
An agent cites a figure that appears in no retrieved document. A model judge rates the answer highly. What should you conclude?
Discussion
Pick a quality standard your team applies by hand today. Work out which of the five rungs could check it: structure, a rule, grounding, a typed decision, or genuine judgement. Most standards people assume need a model turn out to be rules nobody has written down.