TPCN separates a proposed persistent inference system into mechanisms that can be studied independently. Local prediction errors determine a direction of correction; priors constrain states; temporal dynamics set the rate of change; slower estimation updates uncertainty and model parameters.
Approach
The implementation starts with a JAX baseline: state and prediction contracts, several width/depth parameterizations, Euler and Heun solvers, and a synchronous inference runtime. Diagnostics examine equilibrium, termination, and gradient alignment before additional mechanisms are composed.
The broader program includes proximal inference, temporal behavior, representation learning, and eventually active inference. Those are staged research questions rather than features of a completed architecture.
The first decision is whether the baseline behaves consistently as width and depth change. Only then can a new mechanism be evaluated without confusing its effect with a scaling failure in the underlying system. Interfaces remain provisional, and completion of baseline measurements alone does not establish that the full program's integration criteria have passed.