Agentic AI systems increasingly generate, transform, retrieve, and operationalize claims faster than organizations can keep evidentiary evaluation separate from authority to act. This paper proposes a fail-closed epistemic governance architecture in which the complete proposition—not a detached label—is the object under review; evaluation status and operational authority are stored as distinct fields; claim versions are immutable; and new evidence creates successor versions rather than rewriting prior states. The design preserves unresolved but potentially useful candidates without allowing them to acquire authority through repetition, convenience, confidence scores, or later reinterpretation. A binary promotion gate is separated from richer registry semantics, provenance is separated from cryptographic integrity, and formal-verification receipts are explicitly scoped rather than collapsed into a boolean claim of formal verification. The architecture is illustrated using a frozen mixed-domain record containing symbolic lattice candidates, thermodynamic boundary claims, and catastrophic-risk forecasts. Those substantive claims are not reopened. The contribution is an evidence-to-consequence control layer for agentic systems: a deterministic shell that constrains when stored propositions may influence tools, products, or governance decisions. Version 1.1, dated 1 September 2026 (freeze_id: lattice-entropy-gcr-freeze-v1). The PDF was rendered from the author's original .docx (also attached) with LibreOffice; the text is unchanged. The manuscript includes the author's disclosure of generative-AI assistance. Source: author's iCloud folder "Scientific research published final update" (deposited 2026-10-20).
No takes yet. Share an insight, caveat, or question.
Deonte Watts (2026) studied this question.