Background: Machine learning increasingly monitors pharmaceutical manufacturing and quality control. A reader can check such a result only if the paper can be read and reports its data, its code, the data partition and a held-out result. How often both hold here has not been measured.

Methods: Cross-sectional meta-research audit; protocol fixed before screening and scoring. Five bibliographic sources were searched on 23 September 2026 for peer-reviewed studies first published online 2019-2025 that fitted a machine learning model to manufacturing or quality-control data for a human drug or biologic. Outcome 1, among all eligible papers, was access: whether an unaffiliated reader could retrieve the full text without a subscription. Outcome 2, among readable papers, was reporting: nine items from REFORMS, DOME and TRIPOD+AI, and a composite of public data, public code, a stated split and held-out evaluation. A large language model screened and scored under fixed, published prompts, quoting evidence for each judgment; a 20% sample was scored again by a second, independent pass. Proportions carry Wilson 95% confidence intervals (CIs).

Results: Of 9658 records, 3658 were screened, 553 were eligible and 186 were readable (33.6%, 95% CI 29.8-37.7). Among those, 94.0% stated a data split, but 26.7% (20.7-33.6) made data public, 10.3% (6.7-15.5) made code public and 35.1% (28.6-42.3) reported the headline result on held-out data. 2 of 179 met the composite (1.1%, 0.3-4.0). No item differed by venue type. The only year trend below the Holm threshold, uncertainty reporting (p = 0.049), turns on one resolved item and is 0.071 on first-pass scores. All 824 evidence quotes were verbatim in their sources. The two passes agreed on 79.0% of items (kappa 0.64): 94.6% for code, 40.5% (kappa 0.08) for held-out evaluation. Resolving all disagreements moved no estimate more than 0.7 points. No person rescored any paper.

Conclusions: A third of eligible papers could be read without a subscription, and almost none reported what an independent reader needs to re-derive the headline result. Stated splits are nearly universal; public data, code and held-out evaluation are not. Access and checkability are separate gaps; both can be measured at scale.
