What can a generated mismatch dataset or environment contain?

A dataset, task pack or simulated environment for this workflow can contain several layers, and a buyer can ask for any combination.

**Records.** Paired or grouped records from different systems: bank feed lines and ledger entries, purchase orders and invoices, meter reads and bills, claim payments and reserves. Each system has its own identifiers, formats and timing conventions, which is what makes the disagreement believable.

**Relationships.** Links between records, such as a payment to a loan, an invoice to a goods receipt, or a case to the notices it triggered. Some links are clean, some are partial, and some are deliberately absent.

**Policies.** Written tolerances, approval limits, reason-code lists and escalation rules that decide what a role may do. The same discrepancy can have different correct outcomes under different policies.

**Causes.** A labelled root cause for each case, drawn from a taxonomy the buyer can adjust, with decoy explanations that look plausible but fail on checking.

**Tasks and state.** Task statements, the starting state of an environment, and the expected end state.

All of this is generated. Grounding in plausible operating practice is something to inspect in a sample, and the buyer should check it for any resemblance to a specific organisation.