Generated Once, Trusted Forever: Measuring Post-Issuance Drift in AI Bills of Materials

Automated generators now produce an AI Bill of Materials (AIBOM) for a hosted model by reading the model repository's declared metadata. Existing work asks whether such a document can be generated completely; we ask whether the assertions it is built from stay consistent with the repository it read. Reconstructing the retrievable model-card history of 3,000 Hugging Face repositories from their default-branch commit graphs and replaying at every revision the field extraction of a deployed AIBOM generator, we find that 16.8% of the 2,812 models with a usable history withdraw or replace a licence, lineage or training-data assertion after their first model card, and 5.0% do so more than a month later, so this is not only publication churn. For the three AIBOM fields we verify the generator emits it is 16.5%; counting every card-derived registry field, whose narrative sections we show to be an upper bound, 30.7%. In 74.9% of the observable default-branch intervals containing one, no tracked model-artifact file changed, which separates the phenomenon from the familiar decay of a bill of materials whose software has moved. The raw rate rises with popularity, but not within every band of equal edit frequency, and the association attenuates against card-editing opportunity though not against general repository activity. A domain-balanced hand-labelled sample of 72 events finds 80.6% would leave a document that carried the field disagreeing with the repository (kappa = 0.79 against a blind second rater), including 27.8% the raters judged corrections. Simulating issuance over the observed histories, a document issued at a model's first observable card and never revalidated carries a superseded assertion for 15.0% of its life on average over the fields we verify a generator emits, and 28.0% across the registry set; screening for a change costs one request per model for a document that pins its source revision, which one of the two generators we ran does not. We also report two defects and one unimplemented mapping in a generator we exercised.