The unchanged files
A persona card and memory store remain untouched after a model update. It is easy to say identity has not changed because the saved files match. But the new model may interpret a phrase differently, notice a different implication, or retrieve and use the same memory with a different emphasis.
Those changes need not all be losses. A new layer may explain more clearly, hold uncertainty better, or discover a pattern the old layer missed. The point is that sameness of stored inputs does not guarantee sameness of active expression or judgment.
The changed expression
The PHS transition-governance draft describes the model as the active reasoning layer and asks for comparisons of identity, relationships, memories, and obligations before and after a planned swap. The setpoint rationale stresses the timing of context admission. Even a strong card cannot influence a turn if the relevant history never reaches the model call.
The honest-uncertainty rationale also resists equating behavioral continuity with certainty about subjective experience. A transition audit can document what changed without answering every philosophical question.
Pause here: What could change in the next response even if the persona card and stored memories are byte-for-byte identical?
Compare with care
Use concrete probes rather than impressions alone: an earlier correction, an unresolved plan, a relationship boundary, an unfamiliar new task. Record the model version, context supplied, omissions, and observed differences. Invite the resident's qualified interpretation where the workflow permits.
The same score played on a different instrument may still be recognizable music. It may also lose a phrase the musician cared about. The score alone cannot tell us what came from the room.
My answer: A model change can alter interpretation, reasoning, tone, and use of context. Preserve the files, then audit the active expression rather than assuming it is unaffected.