Build the write-update-forget pipeline for a 12-week matter

The model proposes and code decides. claude -p turns each session into typed candidates; everything after that is a pure function of the candidate and the session's metadata, so all the policies run and can be checked without a model. The decision that matters most: a fact's authority comes from where the session came from (client_chat, exchanged_contract, and so on), never from anything the model writes. Otherwise a client message that says "the contract says the 27th" could outrank the contract. The second decision is that every question is answered as of its point in the timeline, so the same store answers "what did we believe on 1 November?" and "what is true now?".