Large language models deployed as persona-conditioned roleplay agents are prone to two related but distinct failure modes: drifting out of character during extended dialogue, and reverting to a generic "helpful assistant" register that displaces the character's own agency. Existing work on persona consistency largely measures whether a model *stays* in character, but treats drift as a single undifferentiated phenomenon and rarely asks whether a model can be *corrected* once drift is diagnosed. We introduce Project Nightcord Sanctuary (PNS), a closed-world multi-agent self-play framework that uses two paired fictional characters as controlled experimental subjects to separate two independent claims: (1) *natural drift resistance* — a model's passive stability in character voice without intervention — and (2) *correction compliance* — a model's ability to actively re-stabilize once a precise diagnosis is supplied. PNS pairs a closed-world container that constrains character knowledge and shared-world facts with a lightweight Router component that scores each turn for drift along two evaluation layers: structural/voice-level cues (language density, sentence rhythm, out-of-character tone markers) that a lightweight judge can score directly, and content-specificity cues (whether an utterance is genuinely character-specific rather than generically plausible) that currently require human domain annotation. We report an illustrative three-round correction case study (v1 raw output, scored as clearly out-of-character; v2 post-Router correction, structurally improved but still hedging, scored 5/10; v3 under an explicit language-density constraint, scored 8/10) that demonstrates measurable correction compliance within a single diagnostic cycle, while the residual two-point gap is attributed to the unresolved content-specificity layer. We position this as a framework paper: the closed-world container, the two-layer Router, and the two-claims decomposition are the primary contributions, with systematic multi-session drift-score data collection ongoing.
Chenge Hu (Fri,) studied this question.