The Independent Oracle Receipt: An Evidence-Dependence Audit for AI-Generated Software Verification

Large language models can generate software implementations and the tests intended to evaluate them. Agreement between these artifacts, however, may be produced by a shared misunderstanding of requirements rather than by independent evidence of correctness. This problem is not new: test-oracle research predates generative artificial intelligence, and recent studies specifically document error propagation, interpretation-coupled testing, and inflated evaluation signals under weak oracles. We therefore do not claim to discover correlated code–test errors or to invent implementation-blind test generation. We propose the Independent Oracle Receipt (IOR) as a prospective, claim-specific method for examining how evidence used to accept AI-generated code acquired its apparent authority. IOR represents a verification episode through a typed dependency graph, separates artifact provenance from semantic warrant, records information exposure and interpretation choices, and subjects material dependency claims to challenge by independent adjudication and controlled exposure interventions. Its output is a bounded evidence-admissibility assessment rather than a certification of correctness, a universal independence score, or permission to deploy software. We define its conceptual objects, auditable states, decision rules, failure modes, and a falsifiable evaluation comparing an integrated IOR protocol with a strong composite baseline that already includes requirement-blind test generation, reference-implementation panels, validity checking, requirement-linked mutation tests, and versioned provenance. The central hypothesis is incremental: dependency-sensitive qualification may reduce unjustified acceptance decisions under matched information and cost conditions. The proposal includes no implementation, experiment, measured benefit, independent replication, proof of universal novelty, or guarantee of software safety.