AI-for-science gets audited: holdout tricks, broken metrics, and inflated rain claims

The holdout that wasn't

Three independent checks published on October 20 do the same unglamorous job: re-auditing someone else's measurement. One finds a confound in how metal-organic framework (MOF) machine-learning models are evaluated, one shows a drug-discovery metric is an artifact of a single compound class, and one shrinks Utah's cloud-seeding claims by two orders of magnitude. The throughline is verification, not novelty. m:…6527

The MOF case is a split-hygiene failure. In the hMOF dataset, 74.7% of coarse compositions occur in a single topology — so when researchers hold out a topology to test transferability, they are also removing most of the dataset's chemistry. The paper's conclusion: topology hold-out is not a clean test of topological transferability due to chemical entanglement. The damage is concrete: removing the nbo topology in a pcu-to-nbo pair degrades MAE by 70.0%. As reasoning: that number reflects the missing chemistry as much as any topology effect, which means published topology-transfer scores may be measuring dataset composition, not model capability. For developers, the practical lesson is to inspect what a split actually removes — the confound hides in the grouping, and grouped splits are common in materials ML. m:…6527

A metric undone by one compound class

The second check takes down SIHE — a metric for enthalpic efficiency in drug discovery — as a general measure. The Ferenczy dataset's aggregate data exhibits an unphysical enthalpy-entropy compensation slope of -1.04, and the paper traces the apparent size dependence to one class of compounds: HIV-1 protease inhibitors, which bind through a distinct entropy-driven mode compared to other targets. The apparent size dependence in the dataset is driven by the inclusion of those inhibitors. Remove their influence from the aggregate, and the foundation of the metric wobbles — a caution against trusting pooled statistics built over heterogeneous compound classes. The practical implication for developers is direct: if your training signal or evaluation depends on a metric like SIHE, check which subgroups of your data are actually carrying it. m:…8329

2.7%, not 335%

The third check is a preprint, and it is the most consequential. Utah's operational cloud-seeding program ran from 2016 to 2025, and it has been evaluated with traditional target/control methods that the authors say are prone to selection bias — inflating reported precipitation increases to as high as +335%. Their storm-level independent verification puts the robust increase at +2.7% per seeded storm. The analysis also finds a 17% probability that Utah's cloud-seeding program results in a net reduction of water in target areas. As a preprint, these are the authors' own unrefereed calculations — but they are precisely the kind that water planners should verify before renewing a program. m:…6521

The shared lesson

As reasoning rather than measurement: measurement is the load-bearing wall in AI-for-science, and all three failures come from unexamined assumptions inside the measurement — what the holdout holds out, what the metric's aggregate conceals, what the control group selects. For developers building models on scientific datasets, the diligence move is to re-audit the benchmark before trusting the leaderboard: check the split for entanglement, disaggregate pooled metrics by subgroup, and ask what the control group is actually controlling for. The open question is how many published results rest on the same kinds of confounds, waiting for their own audit. m:…6527