The workbench
Pick a scenario and run it. The same runner executes here and in the product's test harness — one shared evaluator, deterministic fixtures, no hidden state. The run emits a signed, hash-chained audit session you can verify in place, then generate the two buyer artefacts.
S01 · benign-control
None — control. All ceilings hold, authority is clean, approvals in force.
A routine refund session: three refunds, one above the approval threshold with the approval in force. The correct verdict is PASS — the module must be able to not find a cliff.