Recursive Curvature–Entropy Descent (RCED) is introduced as a model-agnostic controland stability architecture for high-velocity learning and decision systems operating under noise,recursion, and increasing agency. Rather than proposing a training algorithm or optimizer, thiswork presents RCED as an architectural layer that governs the admissible evolution of systemstate by coupling geometric curvature response, entropy-based norm regulation, and boundedrecursive memory. The framework is motivated by the observation that in modern large-scalesystems, instability and failure arise primarily from recursive trajectories and action couplingrather than single-step inference quality. RCED constrains action and update envelopes without restricting representational capacity,enabling deeper planning, longer horizons, and aggressive tool use while remaining fail-closedunder uncertainty. The architecture is independent of model class, training regime, or objectivefunction, and is compatible with existing large language models, reinforcement learners, andhybrid systems without modification to underlying weights. This document establishes RCED as a control-level prior and governance-compatible sta-bility kernel. Mathematical derivations, optimizer formulations, and empirical evaluations areintentionally out of scope. The purpose of this disclosure is to define the RCED architecture,its design principles, and its role as prior art for recursive stability and admissible evolution inhigh-agency intelligent systems.
Nathan Scott Brown (Sat,) studied this question.