Can you keep an AI character in role better by fencing it into a script than by training it?

This explores whether hard-coding what a persona is allowed to do, by confining the model to a structured script, keeps a character steadier over a long conversation than training it with reinforcement learning does. The corpus never runs that contest. The two closest notes measure different things, so 'better' can't be settled from them.

On the RL side, the evidence is a number. Can training user simulators reduce persona drift in dialogue? flips the usual setup and trains the simulated user, not the assistant. It rewards the simulator on three consistency checks: does each line match the persona prompt, does it match earlier lines, and does it survive quiz questions about the character. Drift falls by over 55%. The three checks also separate three kinds of failure: wobble within a turn, slow drift across a conversation, and outright factual contradiction.

The scripting side has no comparable number. Can layered persona architecture sustain coherent character behavior? describes a three-layer persona in which the model only reacts inside a structured script. Its dialogue looks closer to human conversation and is pragmatically fluent, but emotional expression suffers. That note reports human-likeness, not drift reduction, so it doesn't show scripting beats RL at stability. It does show the cost: you get consistency by removing the room to wander, and the room to be expressive goes with it. RL leaves the model free and penalizes wandering. The opposite bet appears in Can personas evolve in real time to match what users actually want?, where personas are deliberately allowed to change at test time. Whether a persona should hold still or adapt depends on the job.

Two other notes suggest the question is missing a third option and a depth problem. How stable is the trained Assistant personality in language models? finds that drift has a direction: emotional and meta-reflective conversations pull models away from the default Assistant along one dominant axis. Capping activations along that axis reduces harmful shifts without hurting capabilities. If drift is that predictable, you may not need a script or a retraining run to control it. Depth matters too. Are RLHF personas performed characters or realized dispositions? and Are LLM personas realized or merely simulated through training? argue that post-training installs dispositions that persist under adversarial pressure, unlike prompt-induced role-play, which collapses. [[persona-steering-operates-in-the-output-channel-and-redistributes-measured-bias]} shows persona prompts changing output without touching underlying bias. That study tested bias, not drift, so it is a hint and not a verdict. But a script that works mostly at the prompt level may hold the surface steady while trained behavior sits underneath, and RL reaches that deeper layer.

The fair test would run both approaches on the same long conversations and score them with the same three consistency checks from the RL study. Until someone does that, the corpus supports a narrower claim: scripting trades expressiveness for steadiness, RL has a measured drift reduction, and steering along the Assistant axis is a third route that neither camp covers.