Beyond the LLM judge: what deterministic checks catch that scoring models miss

Model-graded evaluation became the default because it is easy to set up. It is also non-deterministic, expensive at scale and blind to execution.

Where judges are genuinely good

Tone, empathy, clarity, whether an explanation actually answers the question, these are judgement calls with no ground truth, and a model reads them well enough to be useful.

Where they quietly fail

Ask a judge whether the refund policy was followed and you get an opinion about a policy it inferred from context. Ask whether a tool ran and you get a guess based on the agent's own wording, which is exactly the thing under test.

• Policy compliance: should be a rule engine, not an opinion.
• Tool execution: should be a trace assertion.
• Numeric limits (refund caps, discount ceilings): arithmetic, not inference.
• Required disclosures: string and intent matching against a checklist.

The layered approach

Run deterministic layers first: policy rules, tool assertions, outcome checks. Only send what remains, tone, coherence, helpfulness, to a model.

This makes results reproducible, makes failures explainable by construction, and cuts evaluation cost dramatically because most checks never touch a model.

Reproducibility matters more than you think

If the same conversation scores 0.82 today and 0.71 tomorrow, you cannot use the score as a release gate. Deterministic layers give you a stable signal you can block a deploy on.

What to take from this

• Use models for judgement, code for facts.
• Policy and tool checks must be reproducible to gate releases.
• Layered evaluation is cheaper and far more explainable.