Evals & Review Loops

A golden set is a collection of 20–50 real input→ideal-output pairs, harvested from the manual era of the client's process — before any AI was involved.

Real inputs beat synthetic ones. They surface the edge cases that matter: the weird formatting, the off-topic question, the ambiguous brief that a real user actually sent.

Don't treat the golden set as permanent. Update it quarterly with new failure modes discovered through production sampling. Your eval should get harder over time, not easier.

Build graders in two layers, from cheapest to most expensive.

**Deterministic graders first.** These check structure: is the output valid JSON? Are the required fields present? Is the length within bounds? Are banned phrases absent? Do the links resolve? Are the dates parseable? These run in milliseconds and catch a huge share of failures.

**LLM-as-judge second.** Use this for quality dimensions that require judgment — tone, completeness, accuracy. But anchor it: write a rubric, force 1–5 scoring per dimension, and require cited evidence from the actual output for each score.

Unanchored LLM judges drift optimistic over time. The evidence requirement is what keeps them honest.

For single-step tasks, grading the final answer is enough. For multi-step agents, it isn't.

When an agent runs a sequence of tool calls, you need to grade the entire trajectory: Did it call the right tools? In a sensible order? Within the allotted budget? Without any unauthorized writes?

A perfect-looking final answer can still mask a dangerous or wasteful path to get there.

Extend this logic to the harness itself: test whether the agent can be prompt-injected (feed it adversarial content in retrieved docs and confirm it treats the instructions as data, not commands), test timeout behavior, and check for over-tooling — did it call tools it had no reason to use?

Grade a random 5–10% of live outputs every week. Set alerts for score drift.

Model providers update their models without announcement. Behavior you measured last month may not be behavior you're getting today. Evals are how you notice before the client does.

The weekly one-pager is what turns this into a client-facing asset: outputs sampled, scores by dimension, failures with diagnosis, changes shipped, cost per outcome, next-week bets.

This is the retention mechanism. Clients don't churn from systems whose improvement they watch every week.