What Courts Actually Ask: The Evidentiary Standard for AI-Mediated Harm and Why Governance Documentation Cannot Meet It

AI governance frameworks produce documentation. Courts demand evidence. These are not the same thing. This paper argues that an emerging judicial pattern—visible across Mata v. Avianca (2023), Nippon Life Insurance Company v. OpenAI (2026), and United States v. Heppner (2026)—establishes a de facto evidentiary standard for AI-mediated harm that governance documentation is structurally incapable of satisfying. Courts do not ask whether an organization followed industry best practices or maintained a compliant governance framework. They ask what happened, when it happened, where it happened, who authorized it, and what human principal bears accountability. These are evidentiary questions, not compliance questions. They require a cryptographically anchored, independently verifiable, human-attributed chain of record—not a policy document, a roadmap, or a PDF attestation. This paper identifies the judicial pattern, maps it to the structural gap in current AI governance frameworks, and describes the evidentiary infrastructure that closes the gap: a substrate-layer auditability engine producing GLOBAL-signed, immutable, human-attributed records of what occurred, when, where, and by whose authority—and explicitly declining to confabulate why. The "Never Why" architectural commitment (Truong, 2026, DOI 10.5281/zenodo.19410730) is not a product limitation. It is the honest answer to what courts can actually use.