The Coherence Engine, part deux; Collapsing Bayesian Inference, Free Energy, Gradient Descent, Markov Models, and Reinforcement Learning

Modern AI is driven by five dominant techniques, gradient descent, Bayesian inference, variational free-energy minimization, Markov models, and reinforcement learning. Although historically independent, these methods share a structural assumption; each optimizes a scalar objective inside a fixed representational space. Their successes arise when the world remains compatible with these frozen ontologies; their failures emerge whenever structural novelty is required. This paper reframes these techniques as constrained projections of a deeper coherence dynamic. When the representational space Ω is held rigid, coherence ascent collapses into the familiar behaviors of parameter tuning, posterior updating, or policy optimization. When Ω is allowed to evolve, the full Coherence Engine becomes visible. Adaptive model expansion, structural pruning, and curvature-driven transitions restore the system’s ability to generate new abstractions, causal structures, and strategies. The future of AI depends not on improved optimization within predetermined models but on principled mechanisms for the evolution of models themselves. Coherence dynamics provides that mechanism, unifying inference, prediction, control, and representation under a single meta-law.