C2GRAPH: A Repository-Level Governance Package for Disciplined, Measurable AI Coding Agents

Abstract

AI coding agents increasingly operate as long-horizon software actors: they inspect mutable repositories, call tools, preserve state across interruptions, and may eventually modify artifacts whose failure modes are difficult to reverse. This paper presents C2GRAPH (Convert to Graph Engineering), a repository-side governance architecture that treats these problems as explicit state and authority management rather than as prompt engineering alone. The central design separates evidence acquisition, semantic scope, exact mutation authorization, rollback and freshness, continuity, model/runtime adaptation, and diagnostic recovery. Four extensions—bounded emergency authorization, adaptive context representation, cognitive containment, and progress-liveness self-audit—are deliberately defined as non-authority planes unless they satisfy the same canonical mutation predicates as ordinary work. We formulate this requirement as authority non-interference and evaluate it with process-isolated executable replay, exhaustive enumeration over declared finite abstractions, boundary/equivalence models, and implementation-surface diagnostics. No counterexample was observed in the enumerated authority, liveness, numeric, representation, bounded-read, or adapter domains. The evaluation also discovered two implementation defects that conventional green-path testing had not exposed; both were repaired and converted into regression tests. The results support compositional repository governance within the stated artifact and model boundaries, but do not establish universal correctness, live-provider performance, crash consistency, or human-authentication guarantees.