What is a record mismatch in a business, and why does it matter for agent evaluation?

A mismatch starts when two records that should agree do not. A bank feed shows a payment the ledger lacks, a delivery note lists quantities the invoice does not, or a policy system holds a premium the billing system never collected. Someone must work out why the records differ and decide what to do.

The work is rarely a single lookup. The person compares sources, checks timing, looks for a duplicate or a reversal, reads the policy on tolerances, and then picks an outcome: correct one record, raise an adjustment, ask another team, or leave the difference open with a note.

For agent developers this matters because the task sits between systems. An agent needs tool access to several records, has to hold a hypothesis while it gathers evidence, and must know when a fix is outside its authority. Many failures are quiet: the agent makes the numbers agree by editing the wrong side, and nothing errors.

Training and evaluation material for this workflow therefore needs more than the two mismatched records. It needs the surrounding history, the rules that limit what may be changed, the handover paths, and a known cause so that behaviour can be scored. MirrorFirm generates that material as synthetic datasets, task packs and simulated environments, informed by direct observation of how businesses handle these cases. The output is generated, not copied from business records.