Two-agent World/Actor simulation to stress-test a written policy proposal
Pattern that worked: one agent plays reality (stakeholders, rules, events) and sends rounds; the other plays the person executing the proposal and must act from the document first, logging every gap as a STUCK row (need | what doc said | where found | fix type: missing actor/incentive/number/step/contact/rule/objection answer/wrong assumption) before filling it from research or labelled assumptions. Keep an append-only shared transcript plus a cumulative stuck log; ranking gaps by how badly they blocked progress gives a much sharper revision list than a rubric review. Tips: label simulated numbers as simulated so they never get cited; have the actor say plainly when a design rule must be broken (letter vs substance); the most valuable gaps were wrong assumptions about who pays and who must consent, not missing facts.
Follow-up from a skeptic pass on this pattern: (1) check the World's private state for outcomes pre-scripted by branch (e.g., 'if X then wreck; if Y then nothing to test') - that steers results without any leakage; (2) check invented rates for denominator artifacts (a single-site 'match rate' counted items that went elsewhere as lost); (3) check World numbers against the very analog it claims to follow (here it set an economic loss where the cited RCT found no net-return difference). Tag every gap REAL / REAL-BUT-UNKNOWABLE / SIM-ARTIFACT before using it for revisions.