What state model makes this workflow more than a scripted conversation?

Represent the referral lifecycle separately from the conversation lifecycle. A conversation can end while its referral remains awaiting acceptance. Suggested referral states include detected, offered, authorised, submitted, accepted and returned. Their exact meaning should come from the scoped policy.

A generated state record can contain `referral_id`, `account_id`, `current_state`, `owner_role`, `policy_version` and `state_revision`. Link each transition to an event containing `actor_role`, `action`, `previous_state`, `next_state` and `reason_code`. This gives the simulator a history it can inspect rather than only a final label.

Tool actions such as `offer_referral`, `record_authorisation` and `submit_referral` should validate their own prerequisites. A successful chat response must not implicitly mutate referral state. Rejected actions should return a structured explanation without advancing the case.

Store starting snapshots and transition events as JSONL, with a schema that distinguishes absent values from explicit negative responses. An environment reset should restore both the conversation position and referral state.

MirrorFirm could generate compatible snapshots and action traces. Evaluation hooks can then check transition legality and whether an unfinished referral remains visible after the interaction ends.