What does a sensible two-stage sample review look like for a lab?

An applied research lead at a model lab wants to know whether set-up material suits an agent evaluation. They ask MirrorFirm for a sample covering insurance policy issuance: a bound quote becomes a policy record, a billing schedule and a certificate request.

**Stage one: inspect.** An engineer runs the validation script. Every `quote_id`, `policy_id` and `billing_id` resolves, and statuses follow the permitted transitions. A reviewer with insurance operations experience walks through a handful of tasks with the written underwriting rule. They find one task where the rule is ambiguous and log it as `REV-014`. The privacy lead searches the sample for identifiers and unusual free-text detail, and records the result against the sample reference.

**Stage two: run.** The team loads a slice into their harness, resets the environment per task and runs a current model. They tabulate failures by mode and notice the model often issues the certificate before the first billing record exists.

**Decision.** Because the findings are specific, they ask for a revised sample that fixes `REV-014` and adds more tasks that probe sequencing. They do not yet request volume.

**What it teaches:** a staged review produces concrete feedback to the supplier and a clear basis for the next step. All of the material is generated, so the grounding and disclosure review belongs in stage one.