Courts evaluate a legal-research AI vendor. Which evaluation criterion best reflects AI supply-chain assurance?

Courts evaluating a legal-research AI vendor should require evidence of training-data provenance relevant to that offering. That criterion best reflects AI supply-chain assurance because provenance informs integrity, contamination, copyright/licensing exposure, jurisdiction bias, and suitability for judicial research contexts. Reviewing evidence of output citation provenance relevant to the legal-research AI offering fails conceptually because citation traceability documents sources cited in outputs, not training-data provenance or supply-chain transparency. Verifying evidence of hosting infrastructure provenance for the legal-research AI offering fails because hosting lineage documents where and how the system is deployed, not the origins, integrity, or licensing of the training data that drives model behavior. Evaluating evidence of analyst workflow provenance relevant to the legal-research AI offering fails because workflow history describes analyst handling steps, not the origin or documentation of the training data underlying the AI.

Exam caveat: provenance disclosures may be summarized for proprietary reasons; courts still need enough detail to assess risk classes, not necessarily every raw training row. Operational check: include provenance evidence as a scored evaluation item, reject vendors that provide only analyst workflow provenance, and require contractual commitments to disclose material corpus changes affecting legal-research reliability.