Non-observations are not properties: measurement floors in superconductor databases and their propagation into machine-learning datasets

Superconductor databases record two different kinds of statement in the same field: a measured transition temperature, and the observation that no transition was found down to some temperature. The second is a statement about an experiment, not about a material. Widely used derived datasets encode it as Tc = 0, at which point the distinction is lost. We show that this has two separate consequences. First, in a machine-learning benchmark derived from the SuperCon database, the combination of two individually reasonable cleaning rules — averaging over duplicate records, and discarding a compound when its records scatter by more than 5 K — removes fifteen materials from the dataset entirely, among them Nb3Al (17.5 K), V3Si (16.8 K) and NbN (13.5 K), and shifts the target value of nineteen further entries. A three-line change to the aggregation restores all fifteen and corrects the nineteen without removing any entry or altering the non-superconductor class. The patch has been submitted upstream. Second, and independently, a non-observation carries a measurement floor that derived datasets do not preserve. In MDR SuperCon Ver.240322 we identify 3,261 formulas listed exclusively as non-superconducting; 1,855 of them were never measured below 1.5 K, and 653 have a floor at or above 1 K and no entry after 1980. These are not established non-superconductors; they are compounds untested below the cryogenic limit of their time. We illustrate the second point with a series measured by Ku et al. in 1980, in which 31 compounds of the CeCo3B2 structure type were characterised and 23 recorded as non-superconducting above 1.2 K. Three of those twenty-three have since been confirmed as superconductors — CeRu3B2 at 1.1 K (1985), YRu3B2 at 0.81 K and LuRu3B2 at 0.95 K (2026) — all just below the 1980 measurement floor. Of the three non-magnetic rare-earth members of the series, two have now been confirmed; the third, La0.9Ru3B2, has not been measured since 1980. We report two pre-registered permutation tests that did not support a broader frequency claim, and we state explicitly which conclusions do and do not follow from the data.