Which records can a generated approval dataset contain?

A release for this workflow can be built around a small set of linked record types, so an agent has to follow relationships rather than match keywords.

- **Organisation structure:** roles, reporting lines, business units and the role holders that fill them over time. - **Authority rules:** limits by amount band, risk category, product or asset class, with the role each band requires. - **Delegations:** who delegated what to whom, the scope, start and end dates, and whether sub-delegation is allowed. - **Requests:** the item needing approval, its value, risk flags, requester and supporting documents. - **Decisions and events:** approvals, rejections, returns for information, escalations and reversals, with timestamps. - **Conflict records:** requester-approver overlaps, related-party flags and segregation-of-duties rules.

Policy text can sit alongside the structured rules, including deliberate gaps and places where the written wording and the matrix disagree. Task packs can add the instruction, expected approver, expected outcome and permitted tool actions.

Simulated environments add state: a queue of pending requests, a clock, a roster of who is absent, and downstream systems such as a ledger or purchase order system that change when an approval lands.

Everything is generated. Buyers should ask how the structure was derived, test whether it is plausible for their own process, and inspect the output for any resemblance to material they hold. Nothing in a release should be treated as a record of an actual organisation.