Systemic Epistemic Governance

This paper introduces Systemic Epistemic Governance (SEG), a formal discipline for governing how persistent cognitive systems transform internally derived information into authoritative knowledge and executable action. The central premise is that the ability of a system to derive a proposition should be distinct from its authority to persist that proposition as authoritative or act upon it. SEG formalizes this distinction as a system-wide state-transition invariant. Persistent cognitive state is divided into an authoritative tier, containing externally admitted or independently validated knowledge and executable rules, and a soft tier, containing beliefs, derivations, hypotheses, and generalizations that remain non-authoritative and volatile. The paper provides a small-step operational semantics for SEG and establishes four structural guarantees: authoritative provenance-rootedness, non-circular promotion, execution safety, and soft-tier non-authority and volatility. These guarantees support a separation result termed non-autonomous authoritative escalation: an internal error may propagate through derivation within the soft tier, but cannot become authoritative through derivation alone. Promotion requires a fresh qualifying event supported by evidence independent of the conclusion's own derivation basis. A four-level independence discipline is developed for the promotion boundary: syntactic independence, provenance independence, statistical independence, and adversarial independence. The first is formally proved, the next two are enforced by the evidence construction, and the adversarial level is implemented through an independent confirmation channel and evaluated experimentally. The paper also presents a realization of SEG in a running cognitive architecture, a threat model, and a falsifiable governance-on/governance-off ablation program. Reported experiments examine correlated evidence, adversarial action confirmation, execution safety, and error-cascade behavior. In the reported cascade experiment, governance prevents an injected error from crossing the authority boundary through repeated internal derivation, while the uniform auto-materialization variant allows the error to compound into authoritative claims. The governance experiments were re-run on the live substrate on September 18, 2026. The work positions SEG at the intersection of truth maintenance, data provenance, authorization, information-flow control, and persistent cognitive architectures. Its principal claim is structural rather than statistical: epistemic authority can be treated as a first-class system invariant, making specific classes of unsafe epistemic transitions unreachable by construction. The paper also identifies remaining open problems, including source compromise, provenance forgery, validator error, and adversaries capable of controlling all nominally independent evidence channels.