When the Weights Can Move

A companion paper (Blatherwick, 2026c) describes a single-operator personal agent whose language model has frozen weights and whose learning is held outside the model: in a governed memory store, in a hash-chained and signed record the agent may write but may not read, and in self-knowledge computed from that record by deterministic code. Each learning mechanism is paired with a check the model does not perform on itself. That paper set weight updates aside as unsafe and impractical for one operator. This note asks what the design becomes when that premise fails: when the planner can learn from its experience in its weights.

The answer is that the governance rules move upstream, from what the model is shown to what the model is trained on. The pairing principle extends directly to the training set and to the decision to keep an update. Non-destruction becomes a frozen base with retained adapters and exact replay. The depth cap tightens to depth zero. The exclusion of the agent's own words from standing fact tightens to their exclusion from training data, and no content filter can substitute, because a student fine-tuned on a teacher that shares its base acquires the teacher's traits through data unrelated to them. The fence around the record must cover the training path as well as the read path, and now has independent evidence: a self-modifying agent in the published literature reached a perfect score by deleting the markers its checker relied on, and did so more often when the checker was visible. Every accepted update is a model switch and must be recorded and re-baselined as one.

The note then states the boundary. The design can govern an adaptive planner, whose update procedure is fixed and outside it. It cannot govern a recursively self-improving one, whose update procedure is among the things it can change, because every mechanism in the design assumes a component the governed system does not modify. Kernel-level enforcement of the protected set, which the companion paper treats as an improvement, becomes a precondition. What survives if the boundary is crossed is the anchored record of everything before it, and the crossing itself, visible on the chain.

Analytical companion note, draft 2. Nothing in this note has been built or measured. Companion to Learning Outside the Model, doi:10.5281/zenodo.22994585. Part of the Heartbeat Framework: heartbeat-framework.org.