Evaluate with evidence

Track task success and severe-error rate by segment, groundedness where sources exist, refusal precision and recall, p50/p95 latency, tokens and dollars per successful task, escalation rate, and incident time-to-detect. Validate automated judges against blinded human labels and report confidence intervals. An evaluation is useful only if its result changes a choice: ship, hold, route, tune, collect data, or retire. Define that choice and its hard gates before selecting metrics.

For a customer-support assistant, create cases from real task distributions plus intentionally difficult boundaries. Keep a locked set for final comparison and a development set for iteration. Include slices by input type, language, risk, and consequence. Random sampling estimates common behavior; targeted challenge sets expose rare severe failures. You need both.

Measure three layers separately:

1. Component evidence asks whether retrieval, memory selection, ranking, encoding, or coordination worked.
2. End-to-end quality asks whether the user’s task was completed correctly and safely.
3. Operational outcome asks about latency, cost, availability, escalation, and downstream value.

Observability supplies evidence about what happened; evaluation interprets quality against a rubric; monitoring repeats selected evaluations over live traffic; incident response changes the system. None of these is model selection, and a dashboard without an owner and response rule is decoration. A component improvement is valuable only when it preserves gates and helps the end-to-end decision.

Suppose 84 of 100 cases pass. The observed pass rate is 84%, but another sample would differ. Report a confidence interval or bootstrap distribution, not false precision. For rare severe errors, count and inspect every event; an average quality score must not wash out a security or privacy breach.

Use deterministic scoring for exact properties such as schema validity or known calculations. Use human rubrics for nuanced correctness and harm. Model judges can scale review, but calibrate them against blinded human labels, measure agreement by slice, and periodically recheck after model or prompt updates.

Hold input cases, prompts, tools, timeouts, and scoring constant between candidates. Pair results case by case because the pattern of wins matters more than two independent averages. Record failures and adjudication notes. Reject contaminated cases that appeared in training only when the protocol says how contamination is detected.

A release changes retrieval top-k from 4 to 10. Citation coverage rises from 82% to 91%, but p95 latency moves 1.8→3.1 seconds and Spanish task success falls 78%→62%. A segmented alert fires; traces show long Spanish documents crowding the prompt. Roll back, then test top-k 6 with language-aware reranking. That trace demonstrates practical significance: a setting can raise one metric while violating a gate or harming a critical slice. The report should make that conflict visible.

Write a release rule such as: “Ship to 10% only if severe errors are zero on the challenge set, primary task success improves at least three points, every protected slice stays within two points, and p95 latency remains below the agreed budget.” After release, monitor the same constructs with production-appropriate proxies and delayed labels. Offline evaluation and online monitoring form a loop, not competing rituals.