Bayesian Capability Transport, Disclosure Channels, and Strategic Institutions for Autonomous Agents

This paper develops a formal planning theory for autonomous agents that move across institutions, credential ecosystems, and administrative domains under uncertainty. The core distinction is between three layers: raw institutional states, capability summaries used for planning, and visible summaries used for institution-level evaluation. This separation allows the paper to model what an institution can currently verify versus what an agent must still reason about for future route feasibility, reusable resources, revocation risk, and information gain. On the deterministic side, the paper proves a constructive descent theorem showing when raw transport between institutions induces well-defined transport on capability summaries. It also introduces a stock calculus for consumable receipts, explicitly distinguishing consumption, portable carry-over, and issuance. Because route outcomes are generally path dependent, the positive invariance result is formulated as a normalization theorem: local normalization confluence yields path-independent visible normal forms only relative to a fixed normalization regime. On the stochastic side, the paper rebuilds the framework in Bayesian form. Finite raw transition and observation kernels descend to summary-level kernels under exact quotient conditions, making finite-horizon contingent planning Markov on beliefs over capability summaries. Observation semantics are connected back to the deterministic visible layer through disclosure channels over visible summaries, unifying the deterministic and stochastic parts of the theory. The paper further proves a value-of-information result: Blackwell-more-informative disclosure channels weakly improve the agent’s optimal Bayesian value. The theory is then extended to one-sided strategic institutions that privately know their type and choose garbling channels over visible summaries after the agent acts and selects a disclosure mode. In this setting, a joint belief over capability summaries and institution types is sufficient for finite-horizon planning. Finally, the paper states a minimal exact planning quotient on Bayesian belief states, proves a discounted Bellman contraction theorem for infinite-horizon planning, and gives an approximate-compression bound for deployment-oriented abstractions. Overall, the paper provides a mathematically explicit transport-and-planning layer for autonomous agents that is Bayesian, disclosure-aware, revocation-aware, resource-aware, and strategically aware. It is intended to sit above interoperable credential and presentation standards, supplying the higher-level planning calculus needed for autonomous decision-making across institutions.