ProvelyNo. LLM evaluation scores a model output against a rubric, a dataset, or a judge model. Outcome verification reads the real system after a side effect and decides whether the outcome happened. LLM judgement is never completion evidence.
| Aspect | LLM evaluation | Outcome verification |
|---|---|---|
| Question | Is the output good? | Did the action reach the promised state? |
| Evidence | Text, a rubric, a judge model | Provider state, correlated to the operation |
| Time | Before or after a run, offline | After each action, in production |
| Result | A score | A verdict and a signed receipt |
| Trust in the model | The judge is a model | A model report is E0 and never sufficient |
| Failure it catches | A bad answer | A false completion claim |
When is LLM evaluation the right tool?
Use LLM evaluation for answer quality, tone, and regression testing of prompts. Use outcome verification when an agent makes a side-effecting call and must not say "done" until the provider proves it.
Questions developers ask
Can an LLM judge decide that a refund succeeded?
No. A judge reads text. The refund state lives in Stripe. Provely reads Stripe with a read-only credential and compares the state with the contract.
Does the Integration Compiler use an LLM?
Yes, for semantic analysis and contract synthesis, bounded by provenance and conformance gates. Every semantic claim cites a source. The runtime uses no LLM.