Agency and Unification under Epistemic Uncertainty

The thermodynamics of prediction relates dissipation to \emph{nonpredictive memory} (\emph{nostalgia}) in driven stochastic systems \citep{Still2012ThermoPrediction}. This paper develops a clean epistemic-only'' specialization aimed at real physical settings and embodied agents: underlying world dynamics are deterministic, and all apparent stochasticity arises from uncertainty about initial conditions (and, optionally, about fixed law bits'') together with many-to-one sensory coarsening. We model sensing and actuation \emph{interfaces} as \emph{bridge variables}: each has an agent-owned component (part of the agent state, directly reconfigurable) and an environment-owned component (channel conditions, treated as part of the hidden microstate), thereby modulating the effective sensing/actuation maps. We define plasticity using directed information \citep{Massey1990DI,Abel2025Plasticity} and define empowerment as state-conditioned channel capacities (equivalently, reachable-set size in deterministic dynamics) under the epistemic trajectory distribution induced by the prior over latents. A central technical contribution is H1: \textbf{perfect prediction/compression is achieved by \emph{identifying} the latent microstate (Route~1) or by \emph{overwrite control} that makes future trajectories action-determined (Route~2)}; empowerment (as an action$\to$world capability measure) is generally \emph{not} sufficient for perfect prediction. We prove that (i) maximizing prediction/compression of observations drives hidden-state compression when interfaces are refinable and agent memory is sufficient, and (ii) whenever refinability requires steering world-side interface conditions, prediction pressure implies nontrivial \emph{interface empowerment}. We then record a bit-string ``agent=human, environment=universe'' specialization with a conserved information budget, yielding explicit saturation trade-offs between prediction-by-identification and control-by-overwrite.\footnote{This is a working draft. Feedback and criticism is most welcome.