How does a mismatch appear in the systems an agent connects to?

In a connected process, a mismatch is a relationship between records in different systems, not a field in one. An agent sees it only if the tools expose both sides and the link between them.

A typical shape has: - a **source record**, such as `BANKTXN-5521` in a bank feed; - a **target record**, such as `LEDGER-88140` in the ledger; - a **match key**, for example a payment reference, that is sometimes missing or malformed; - a **variance** with amount, direction and age; - a **case** record, `CASE-0412`, holding status, owner and notes.

The tools an agent usually needs are read actions (`get_record`, `search_records`, `get_history`), a policy lookup, and a small set of write actions (`propose_adjustment`, `post_correction`, `reassign_case`, `add_note`). Write actions should return structured errors when permissions or approvals block them, because that behaviour is part of what you want to test.

Time matters too. Records carry created, posted and reversed timestamps, so an agent can tell a timing difference from a true error.

Ask for the case record to link back to the generated root cause under a separate key that the agent's tools cannot read. That keeps scoring possible without leaking the answer into the agent's context.