AI characters

Problem. The model is told everything and then asked to pretend it does not know. It leaks. Retrieval returns whatever embeds close to the query, including the twist.

How it maps.

• context.admit admits candidates for one perspective within a budget and returns a reason for every refusal: future, inactive, not-held, withheld, excluded-classification, duplicate, over-budget. Nothing is dropped silently, so when the model gets something wrong you can see whether the context was at fault.
• A character-voiced generation uses that character’s perspective; a narrator has none and sees everything visible. The distinction is enforced, not prompted.
• Model output is a proposal, not canon: review per operation, apply what you accept, keep the rationale of what you reject. If the story moved on meanwhile, proposal.rebase says which operations still fit.
• A generated scene can also be written to a draft and promoted after review.
• Finding what is relevant is a search by meaning: the client’s embedding model turns texts into vectors, Umwelt stores them and searches them within the same perspective. “Who took the necklace?” finds “someone stole the duchess’s jewel” in what this character holds, never a twist they have not learned. See ../spec/search.md.

Umwelt treats generated content as something to review rather than something to trust, using the same machinery a human editor would use.