The Half-Life of an Open Model: An Empirical Audit of Gating, Deletion, and Licensing Mutation in the Machine Learning Commons

The Hugging Face Hub has become the default distribution layer for open machine learning artifacts and, in the process, a de facto citation target in the empirical software engineering and natural language processing literature. Prior ecosystem studies characterise the Hub cross-sectionally: they count what is present. We invert the question and measure what disappears from the scholarly demand side. We mine the full text of arXiv (January 2020–June 2026), extract every Hugging Face model and dataset identifier referenced in scholarly work, and resolve each identifier against the Hub metadata API through an authenticated crawl to classify its present state within the space S = {Active, Gated, Deleted, Renamed, Mutated}. Authentication is essential: it separates genuine deletions (a clean HTTP 404) from access gating (the gated flag), a distinction the Hub collapses into an ambiguous 401 for unauthenticated callers. On a date-windowed 2020–2026 cohort of 307 confirmed identifiers (after removing Hub-UI false positives) cited across 160 papers, only 31.3% remain Active in their original, freely retrievable form: 32.6% return a genuine HTTP 404 (deleted), 29.0% have been renamed to a successor namespace, and 7.2% sit behind authorization gates—so 61.6% of cited artifacts are no longer served at the identifier the literature records. Fitting Kaplan– Meier availability curves over the multi-year cohort, the empirical availability half-life of a cited open model is 10.1 months. A Cox proportional hazards model (corporate reference) isolates the mechanism: conditional on artifact size, independent single- user provenance carries 1.84× the deletion hazard of corporate provenance (95% CI 1.22–2.79), while gating is a near-exclusively corporate behaviour (13.7% of corporate artifacts versus 0.0% of independent). Two quantities remain deferred and we say so plainly: silent license mutation (the Internet Archive Wayback endpoint was unreachable during the crawl) and exact deletion timing (the Hub exposes no deletion timestamp, so event times are interval-censored). We argue that a reference to a mutable commercial hub is not a durable scientific citation, and propose content-addressed archival, mandatory weight DOIs, and license- drift monitoring as concrete mitigations.