Access Is Not Yet Verifiability: Toward a Claim-Preserving Evidence Contract for AI Assurance

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Access to an AI system is a prerequisite for meaningful evaluation, but access alone does not make an assurance claim verifiable. This position paper examines the gap between evaluator access and evidence that remains interpretable and reviewable outside the original evaluation context. It proposes a claim-preserving evidence contract for AI assurance: evidence should preserve the connection between the claim being evaluated, the relevant evaluation context, the observed result, its provenance and scope, and the limitations on what that evidence can establish. The paper distinguishes access, evidence, verification, and assurance, and argues against treating availability, signatures, logs, or evaluator status as substitutes for claim-level evidence. This is a non-normative Phionyx Research position paper. It does not modify the AIREP specification and does not claim certification, accreditation, evaluator independence, model safety, or production assurance.