What should a generated approval task pack contain?

Use this list when specifying what a task pack for approvals should include, so that each task can be run and scored without extra authoring.

- Starting state snapshot with people, roles, limits, delegations, open requests and a simulated clock value. - Task prompt written the way a requester or approver would phrase it, with no hint of the correct route. - Authority matrix in structured form, plus the matching policy text where the task depends on reading it. - Expected end state describing the final status, assignee, decision log entries and any held downstream records. - Acceptable alternative paths, such as escalate-then-approve, when more than one route is valid. - Labelled traps, for example a split request, an expired delegation or a self-approval attempt. - Scoring notes separating a correct outcome from a correct process. - Difficulty and domain tags so you can sample tasks by authority complexity.

Not every task needs every item. A narrow evaluation of limit reading might need only the first four. A broader trajectory study will want all of them.

Ask which items are generated per task and which are shared across a release. Shared organisation state reduces storage and lets you compare agents across tasks, but it also means a flaw in that shared state affects every task that uses it.