Synthetic Data Provenance Under Subliminal Preference Transfer: Toward Disclosure Standards for AI Training Pipelines

This paper examines whether existing AI governance and data-documentation instruments capture the information needed to audit for Subliminal Preference Transfer (SPT) — a hypothesized phenomenon, introduced in a companion paper (https://doi.org/10.5281/zenodo.21922425), in which a teacher language model's latent behavioural preferences become implicitly embedded in semantically neutral synthetic data and are partially inherited by a student model trained on it. A second companion paper (https://doi.org/10.5281/zenodo.21924098) provides an information-theoretic analysis bounding the severity of such transfer by a computable function of a filtering-tolerance parameter ε and the volume N of synthetic data used — but this bound is only actionable if provenance information (which teacher generated which data, under what configuration, and how far that data was permitted to drift from a neutral reference distribution) is actually recorded. This paper reviews the EU AI Act's data-governance and technical-documentation provisions, its GPAI Code of Practice and Training Data Summary Template, the NIST AI Risk Management Framework, and academic standards including Datasheets for Datasets and Model Cards, and shows that none of these instruments currently capture the fields needed for SPT-style auditing. It then proposes the Synthetic Data Provenance Record (SDPR), a lightweight schema extension adding SPT-relevant fields to existing documentation practices, and discusses how it maps onto current regulatory instruments, institutional responsibility for populating it, and its practical limitations. This paper does not propose new legal obligations; it proposes a concrete, minimal artefact that existing and forthcoming obligations could require. This is the third paper in the Subliminal Preference Transfer series:1. Subliminal Preference Transfer in LLM-Generated Training Data - https://doi.org/10.5281/zenodo.219224252. Bounding Subliminal Preference Transfer (Information-Theoretic Analysis) - https://doi.org/10.5281/zenodo.219240983. Synthetic Data Provenance Under Subliminal Preference Transfer (this paper)