Verifiable Longitudinal AI Records: Provenance as a Methodological Requirement for Human-AI Trajectory Research

Longitudinal evaluation is becoming increasingly important as AI systems persist across longer tasks, retain memory, and participate in extended human-AI relationships. Yet trajectory- level analysis usually assumes that the underlying interaction record is trustworthy. This paper argues that this assumption should be made explicit and testable. A longitudinal evaluator can only assess the trajectory it is given; if that trajectory has been edited, selectively omitted, summarized, reordered, detached from its model or context, or altered after the fact, sophisticated analysis may still rest on an unreliable measurement substrate. We distinguish content validity - whether an AI output is true, safe, appropriate, or useful - from record validity - whether the evaluated record is the authentic, ordered, and unchanged record of what occurred. We propose a provenance architecture for longitudinal AI research built from content-addressed hashing, cryptographic timestamping, immutable public provenance, private-state separation, version and correction lineage, observable execution-context binding, and portable verification packages. We then describe an implementation path spanning Memory Chain, a provenance bridge to an adapted open-source Mempalace memory layer, and Haawke Phoenix. Engineering acceptance tests in Phoenix demonstrate provenance-preserving deletion: private artifacts can be deleted while their public cryptographic records remain independently verifiable, and exported provenance packages can be checked without exposing private system instructions or account/contact fields. The paper does not claim that provenance establishes factual truth, full reproducibility of opaque provider-controlled model state, or model identity in a metaphysical sense. Its narrower claim is methodological: trajectory-level research should treat provenance of the trajectory record as part of the measurement model.