Define what good means
Choose task-specific measures: correct classification, supported answers, safe tool arguments, response time or completed work accepted by a reviewer. State the denominator. An accepted task and a model response are different units; hiding that distinction makes cost comparisons misleading.
Test the system, not only the model
Include retrieval, tools, approvals and recovery in the test boundary. Keep representative examples, known failure cases and an agreed release threshold. Record the version and configuration used for each run. A useful evaluation can be repeated after a change.
Use observation to improve
Production traces can reveal unexpected paths, but their collection and retention need permission and minimisation. Review incidents and regressions against the original task contract. This site provides a planning checklist; it does not run a benchmark, validate a model or certify a deployment.