What can a generated mismatch release contain?

A generated mismatch release can contain the pair of records, the surrounding context and the correct outcome. The point is that a model or tester can be scored on the whole chain, not just on spotting a difference.

Typical contents:

- **Record pairs:** ledger entries against bank lines, invoices against goods receipts, meter reads against bills, stock counts against inventory balances. - **Relationships:** shared references, customer or account links, batch identifiers and the timing gaps between the two sides. - **Cause labels:** timing difference, duplicate posting, keying error, rounding, currency conversion, missing handover or unmatched write-off. - **Policies:** tolerance bands, approval limits, rules on reversal versus overwrite, and escalation paths by role. - **Tasks:** find the cause, propose the correction, draft the adjustment, or decide to escalate. - **Environment state:** open and resolved items, queue assignments, audit logs and the permissions each role holds.

A dataset holds the static cases. A task pack adds instructions and expected outcomes. A simulated environment adds tools that read and change the records, so an agent's actions alter state that can be checked afterwards.

Cases can include the awkward ones: two causes at once, a correct record that only looks wrong, and items that no permitted fix can resolve. Scope the mix in a pilot, and inspect how well the generated cases hold together.