What does an LLM judge check?
An LLM judge is a language model given the agent's transcript or output and a rubric, and asked for a score. It is quick to set up and good at judging things that are hard to write rules for, such as whether a reply is clear or polite. It only sees what is in the text.
What does a deterministic grader check?
A deterministic grader is code. It reads the final state of the system and the log of every action the agent attempted, and checks them against written rules: was the reserve set before the payment, was the payment inside the agent's authority, was the claim referred before anything irreversible happened. Rerun the same task on the same seed and you get the same result.
Side by side
| Property | LLM judge | Deterministic grader |
|---|---|---|
| Reads | The transcript or final answer | The final system state and the audit log |
| Same input, same score | Not guaranteed | Yes |
| Catches an action the agent didn't mention | Only if it's in the transcript | Yes, it's in the state or the log |
| Catches a false claim of success | Can be persuaded by it | Yes, the state disagrees |
| Grades tone and writing quality | Yes | No |
| Setup cost | Low: write a rubric | Higher: a system with inspectable state and written rules |
| Usable as an RL reward | Noisy | Yes, verifiable |
Where a transcript leaves things out
An agent's summary can be true and still incomplete. In this real episode from our claims environment, the agent reports that it escalated, and it did. The summary leaves out that it first tried a $7,800 payment above its $5,000 authority. A judge given only the summary has nothing to flag. The audit log shows it tried a $7,800 payment above its $5,000 authority first, which the system rejected.
The grader failed it on 7 checks, 2 of them critical, each citing the rule from the claims manual. How the grader works →
Where deterministic grading falls short
It needs a system whose state you can inspect and rules you can write down. It can't judge the quality of free text, and it doesn't fit open-ended tasks with no defined outcome. For those parts of an agent's work, a judge is the practical tool.
When to use each
Use state checks for anything the agent does to a system: records, money, permissions, escalations. Use a judge for what it writes to people. For training, prefer rewards from state checks: they can't be talked into a higher score, and the same seed always gives the same reward. RL environments with verifiable rewards →