Experience Sustaining: A Systems Framework for Adaptive Inference in Human–AI Interaction Toward Efficient, Ethical, and Deep Human–AI Co-Evolution

Current large language models optimize for task completion, systematically eroding the semantic-cognitive continuity required for open-ended inquiry, ethical deliberation, and long-horizon learning. We propose Experience Sustaining (ES), a runtime interaction framework designed to preserve Semantic-Cognitive State Continuity (SCSC) — the coupled system property that keeps the human cognitively engaged, the interaction semantically diverse, and the conversation resistant to premature resolution. The framework formalizes SCSC as the primary conserved variable; a four-dimensional interaction space (Novelty, Coherence, Cognitive Effort, Directionality — N-C-E-D) with computable proxies; a discrete intra-regime stability metric (Index of Sustained Experience, ISE); four collapse conditions; and a model-agnostic runtime policy layer deployable on existing LLMs without retraining.

This version introduces the Differential Collapse Postulate: the hypothesis that distinct task types produce structurally different SCSC collapse trajectories. In high-stakes deliberation tasks, collapse is driven exclusively by Novelty Exhaustion while human effort remains elevated. In divergent creative tasks, Effort Abandonment precedes Novelty Exhaustion by approximately two turns. In structured learning tasks, Novelty Exhaustion precedes Effort Abandonment by approximately three turns. Each regime requires a different primary corrective intervention. An agent-based simulation model (n=200 conversations per condition; seed=42; 15 turns; three policy conditions: Baseline, ES-lite, ES-full; four task types) generates results consistent with the postulate. Collapse rates of 58-100% under Baseline in open-ended tasks are plausibly reduced to 0% under ES-full. All results are proof-of-concept under simulated conditions; human-subject validation is required.

Supplementary materials include the simulation model (es_simulation_v3.py, Python/numpy, fully reproducible) and complete results (simulation_results_v3.json).