Federated Epistemic-Operational Architectures for Evidence-Governed AI Systems

Abstract
Modern AI systems often combine hypothesis generation, uncertainty handling, evidence acquisition, and goal-directed decision-making within the same subsystem. While this integration can produce highly capable systems, it also creates fundamental difficulties in transparency, verification, and governance. When a system that generates a hypothesis is also responsible for evaluating whether that hypothesis is sufficiently supported, and for deciding what actions to take on its basis, the epistemic and operational functions become entangled in ways that are difficult to audit, correct, or govern at scale.
This paper proposes a Federated Epistemic-Operational Architecture (FEOA) that separates these concerns into distinct layers, grounded in three constitutional design principles: that hypotheses must remain revisable, that no subsystem may assume complete knowledge or deterministic control, and that cooperative evidence exchange between constitutionally aligned agents measurably improves epistemic quality and adaptive capability. These principles are stated as engineering requirements — empirically supportable claims about what well-functioning AI systems operating under uncertainty must do — whose consequences determine the architecture.
The architecture consists of three layers: an Epistemic Layer responsible for maintaining hypotheses, evidentiary support structures, confidence assessments, and contradiction detection; an Epistemic-Serving Operational Layer responsible solely for uncertainty reduction and evidence acquisition on behalf of the Epistemic Layer; and a Goal-Directed Operational Layer responsible for external action governed by governance constraints. Multiple agents may participate in a federation while preserving local epistemic sovereignty through provenance-aware evidence exchange, where trust governs interaction allocation rather than belief formation.
The architecture is model-agnostic and may supervise predictive systems, generative systems, retrieval systems, reasoning systems, and autonomous agents. Rather than replacing existing AI models, the proposed framework treats them as sources of hypotheses whose outputs are evaluated through an explicit epistemic process before operational adoption. One concrete Layer 1 implementation demonstrating the feasibility of the Epistemic Layer is described in a companion paper.