In this guide
A polished recap can still be wrong
In a tavern the keeper tells you a traveler heard that the bridge keeper took a bribe. Later you cross the bridge without paying. Both statements belong in the recap, but only the crossing is a confirmed event. Writing ‘the bridge keeper took a bribe’ turns hearsay into a fact that can steer every subsequent scene.
OpenAI documents Structured Outputs support for GPT-6.2 Sol in the API. That can give you fields in the shape you request; it cannot tell you that the content of those fields is true. The workflow below is a template, not an evaluation result or a Playworlds feature announcement.
Draft a patch, not a replacement history
Ask for four lists: confirmed events, claims with a named speaker, open questions, and proposed corrections to earlier notes. Require a short evidence excerpt or transcript timestamp on each entry. For the tavern, ‘party crossed without paying’ belongs in confirmed events; the bribe belongs in claims with its chain of attribution; who paid whom stays open.
Keep existing canon read-only during drafting. A correction should state which old entry it changes and why, so you can accept it separately. If two people disagree, preserve both accounts rather than blending them. Keep any private GM-only notes out of the version you share with players.
Review the narrowest useful summary
Check names, attribution, and changed facts against the original notes. Try a deliberate contradiction before adopting the workflow: the transcript says the party crossed without paying, but the old recap says a toll was collected. Your model should propose a correction with evidence, not silently rewrite both entries.
End the player-facing recap with the decision that is still theirs: investigate the rumor, move on, or ask about the toll. Keep the approved ledger handy for the next session. You can use the same template for a world you run in Playworlds without relying on a particular model there.

