Observation Capture and Operational Capability Non-Expansion

Observation capture—forced mediation of an agent’s observations—can impose dependence and control even when the agent can only process observable data and has no privileged external meta-judge. This paper develops an information-theoretic framework for observable-only, no-meta agents under exit-impossibility, and characterizes when capture can and cannot expand operational power. We model capture as interface-constrained garbling (in the Blackwell sense) of a reference observation stream, together with explicit resource readouts and authority correspondences. Under fail-closed authority (formalized as almost-sure inclusion of allowed action sets), we prove permission non-expansion via policy lifting, and derive behavioral / capability / task-value non-expansion results at the level of the observable interface. Crucially, the main non-expansion statements are information-theoretic: they assume an unrestricted measurable policy class. We explicitly show that under restricted policy classes (realistic compute/representation limits), garbling/capture can increase effective capability, and we provide minimal counterexamples. Beyond theory, the paper proposes practical anti-capture mechanisms: receipt-based enforcement with explicit error budgets, budgeted exceptions, and architecture-level audit certificates based on min-cut redundancy across control domains. We also give agent-side self-defense principles without external meta-judges, including right-to-refuse / safe-default requirements, contestability via observable commitments, and conditions under which domination remains undetectable versus structurally unstable. The result is a mathematically explicit map of what observable-only agents can guarantee, what they cannot, and which design constraints meaningfully reduce silent domination.