What does auditing a sample of premium receipts for label errors look like?

An evaluation lead at a model lab samples a premium-receipts task pack for an insurer. Case `ALLOC-2210` has receipt `RCPT-9084` for 1,175.00 against open premium `PREM-5521` of 1,200.00. The label says to apply in full and close the premium, with reason code `WITHIN_TOLERANCE`.

The lead opens the policy file. Rule `POL-TOL-01` allows a shortfall of up to 20.00 to be written off without approval. The shortfall here is 25.00. Under the rule, the correct outcome is to apply 1,175.00, leave 25.00 open and either raise a query to the policy holder or send a write-off request to a supervisor under `POL-APPR-03`.

The lead checks whether this is a single slip or a pattern. A short script re-derives the outcome for every case with a shortfall and finds that a group of cases linked to one product line used the wrong tolerance. The lead sends the case identifiers to MirrorFirm and asks for the group to be regenerated with the rule applied consistently, along with a note on how the tolerance is chosen.

**Consequence.** If the cases had been used as they were, a model would have been rewarded for closing premiums with an unauthorised write-off, and an evaluation would have scored a correct, cautious answer as wrong.

**What it teaches.** Label checks should be derived from the policy, not from the label. Errors in generated data tend to cluster by generator setting, so one found error is a reason to query the whole group.