Epistemic Architecture: A Constraints‑and‑Objectives Framework for Bounded Agents Under Adversarial Pressure

What system emerges when agents must act under uncertainty, coordinate with others, and survive errors—all within bounded verification capacity and under adversarial pressure? This paper derives an answer from minimal constraints on agent and world. The central move is a type‑separation: ordinary epistemic vocabulary conflates two operationally distinct phenomena. Traction is internal action‑guiding grip— private, cheap, revisable. Authorization is ledger‑mediated permission to rely—gated, process‑mediated, scoped. This distinction immediately reframes canonical puzzles: Gettier cases are provenance‑field failures (the justification chain doesn’t track the actual truth‑maker—not a missing fourth condition); Lottery cases are type‑confusion between traction and authorization. Each puzzle localizes to a specific architectural component rather than demanding a new analytical clause. The architecture that makes these localizations precise comprises scoped trust zones (Bubbles), a shared authorization ledger (Bank) with lifecycle gates (submit, challenge, quarantine, repair, revoke, export), structured validation headers encoding standard, error model, and provenance (S/E/V), and temporal validity constraints. These primitives are not stipulated but pressured: given bounded verification, deception, staleness, and coordination need, alternatives fail in characteristic ways. Convergence across independent domains—science, law, network security, supply chains— provides diagnostic evidence for the constraint bundle, which identifies its own falsification conditions. A machine‑checked Lean companion (435 theorems, zero sorries) proves the conditional necessity results, provides non‑vacuity witnesses, and verifies that compatible extensions cannot invalidate established properties.