LLM × User Profiles

Central question. A profile is a lossy, updateable interface between behavior and a model, not a true description of the user. A complete design must specify what evidence is compressed, when an event becomes durable memory, how the consumer uses the profile, how unsupported claims are detected, and how the user can correct or refuse it. The roadmap includes a full faithfulness and governance section because downstream accuracy alone cannot answer those questions.

Checklist before deployment. For every profile, ask: which events support each claim; what is transient; what writes persist; which model consumes the result; what errors the evaluation counts; and whether the user can inspect, edit, delete, or opt out. If these answers are missing, a fluent profile can still be an ungrounded and hard-to-correct guess.