Beyond Knowledge Editing: A Coherence-Governance Framework for Long-Horizon Multi-Module Industrial AI Agents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Continual-learning industrial AI agents — multi-module systems combining perception, symbolic representation, planning, tool use, and uncertainty monitoring — accumulate inter-module representational drift over their operational lifetime. Building on recent work on knowledge editing in continual-learning systems, this paper argues that local knowledge editing, while necessary, is insufficient for the class of failures that emerges in long-horizon multi-module deployments, because it operates at a granularity finer than the unit at which coherence collapses occur. We propose a governance framework that goes beyond knowledge editing: the cognitive stack of an industrial agent is decomposed into six representational layers, and inter-layer coherence — rather than the correctness of any individual fact — is treated as the primary variable to be monitored and governed. We define a Global Cognitive Resonance Index R(t) as a measurable function of inter-module representational alignment, computable in real time from quantities already available in standard agent architectures (embeddings, output distributions, probe-set consistency), and an Ontogenetic Sculpting Operator S as a governable, auditable re-alignment policy triggered when R(t) falls below a critical threshold for a sustained interval. We instantiate the framework on a reference industrial scenario — a multi-module decision-support system for predictive maintenance under continuous learning — and provide an empirical proof-of-concept on the public NASA C-MAPSS benchmark, in which the six layers are instantiated with real components and a continual-learning schedule induces genuine inter-module drift. Averaged over multiple seeds, post-hoc interface re-calibration under S recovers task performance that a local-editing baseline does not — local editing in fact degrades it — while an experience-replay baseline that mitigates the drift during training is stronger still; we therefore position S as a post-hoc governance intervention for the setting in which the upstream module cannot be retrained with rehearsal. R(t) proves an effective drift monitor and trigger but an imperfect proxy for task performance. Recent empirical work on shutdown resistance and multi-agent shut- down sabotage sharpens an adjacent architectural point: prompt-level compliance and locally well-behaved components do not by themselves guarantee system-level controllability. We therefore explicitly separate coherence governance from a non-bypassable external interruptibility boundary whose authority is not writable by the governed agent or its peers. We do not yet report results on a deployed system. The design is informed by the way biological cognitive systems maintain coherence across hierarchically organised scales, but this biological motivation is treated strictly as design inspiration and not as explanatory foundation. Speculative theoretical extensions and candidate hardware substrates that motivated the present formulation are isolated in a dedicated Outlook, so that the operational core can be evaluated independently.