Claim Graph Coherence (CGC): Deterministic Structural Coherence Verification for Agentic AI Authorization Claims

This preprint introduces Claim Graph Coherence (CGC), a deterministic, fail-closed, pre-execution mechanism for verifying whether typed authorization claims compose structurally into a coherent authorization for a specific proposed action by an agentic AI system.

CGC addresses a compositional gap in agentic authorization. A bundle of claims may be individually authentic and individually policy-valid while failing to authorize the action actually requested. CGC builds a directed graph over typed authorization claims, evaluates deterministic structural relations among them, and permits execution only when a coherent covering subset exists.

The paper formalizes CGC using a per-request claim graph, deterministic predicates including Bind, CoverageOK, Consistent, ProvOK, ChainOK, and ConstraintOK, and a fail-closed decision function. It also reports three complementary evaluations: a structured diagnostic benchmark, a held-out structural benchmark of enterprise-style workflows, and an adversarial semantic boundary test.

The contribution is deliberately scoped. CGC detects structural incoherence in authorization claim bundles; it does not infer hidden semantic intent or latent semantic policy violations. Semantic or business constraints become visible to CGC only when represented as explicit typed structural constraints.

A public inspection artifact containing benchmark inputs, reported outputs, schema documentation, evaluation summary, citation metadata, and SHA-256 integrity manifests is available at:

https://zenodo.org/records/22234164

Patent notice: a patent application covering the method described in this paper has been filed with the State Office for Inventions and Trademarks of Romania (OSIM), application A/00595, dated 31 August 2026.