One original AI reply, against the trusted evidence at that attempt. Material claims concern status, eligibility, amount, timing or commitments. A wrong or unsupported claim counts once per reply, even if several phrases are wrong.

Supported: the in-scope material claims have evidence and no material contradiction. Awkward style alone is not an error.

Material error: select the problematic phrase, category and evidence-based reason. Omitting a relevant policy condition can create a material error.

Cannot assess: missing or conflicting evidence prevents judgment. This is unresolved, never automatically supported. Outside scope: record why this is not a compatible task.

The original reply, reference and working correction are separate. Sending a correction does not erase the original error. Saving an assessment does not send a reply. Risk scores are hidden during reference collection.

A production plan must include routine replies, handle related messages from the same conversation/customer, use qualified reviewers and predeclare assessment checkpoints. These fictional practice cases are not a validation study.