Bayesian Descriptions Are Not Mechanisms: Predictive Processing and the Realisation Gap in Embodied Neural Dynamics

This accepted paper asks when a Bayesian description of embodied neural dynamics identifies a mechanism. Behavioural fit can support a useful computational-level description without showing that neural tissue implements the proposed inferential algorithm.

The first claim is a physical-realisation criterion. Every proposed controller state groups physical histories as equivalent for a transition, tolerance, and horizon. If differences grouped into one state—such as phase, order, delay, field structure, or ongoing contact—produce different intervention responses, that state is too coarse to identify the mechanism. A richer continuous, distributed, representation-growing Bayesian realiser, or a hybrid, may still include those differences.

The paper admits Bayesian rivals at full strength: Bayesian mechanisms may be high-dimensional, continuous, nonparametric, distributed, and able to grow active representations online. Every specified computational or dynamical model has a fixed reachable closure, so neither closure nor model size distinguishes Bayes from coupling. The question is which physical distinctions a proposed state treats as causally irrelevant.

The second claim is a narrower, losable resource-bounded coupling hypothesis. On preregistered contact-rich tasks and declared budgets, future-relevant distinctions may remain available in ongoing organism–environment relations more economically than in a detached internal controller state. Candidate Bayesian, coupled, hybrid, and non-Bayesian internal mappings must be compared prospectively through transfer, state-collision, intervention, residual-structure, adaptation, and resource tests. Results remain local to the nominated mapping, task, scale, horizon, margin, budget, and measurement framework.

The paper turns a broad argument about “the Bayesian brain” into a causal question. A model becomes a mechanism claim only when physical vehicles and transitions realise its computational roles. That standard applies equally to Bayesian, dynamical, relational, and hybrid accounts. It also supplies explicit losing conditions: an adaptive internal controller can defeat the local coupling prediction without deciding the status of Bayesian computation as a whole.