When Is a Bayesian Posterior Actionable? Runtime Admissibility under Distribution Shift Reality-State Typing and Gating Layer (RSTG-L)

Bayesian Deep Learning (BDL) provides calibrated predictive uncertainty, yet in deployment the central question is prior to calibration: when is a Bayesian posterior admissible to drive action? Under distribution shift, feedback, and regime transitions, probability and utility can become ill-posed because the operational reference class (the present mechanism regime) is not fixed. We introduce Reality-State Typing and Runtime Admissibility Gating, a keeper-first decision layer that anchors a present reality-state type, validates regime consistency, and gates mechanisms and interventions via a minimal control tuple. “Coordinate failure” is treated as a robustness requirement: when decidability cannot be established the system degrades to a bottom state, restricts actions to a safe action set, and enforces “no escalation by default.” Only when validation remains stable for T consecutive steps are downstream probabilistic and utility components admitted in typed form, enabling planning rather than passive prediction. We provide a falsification-first evaluation protocol with stress tests and auditable metrics for invariance, gate compliance, degradation, and recovery. The proposed layer does not replace BDL or causal world modeling; it specifies a deployable admissibility interface that makes posteriors and utilities eligible to act under explicit regime typing.