Durability of United States Tariff Classification Rulings, 1989 to 2026: data, code and analysis plan

Data, code and analysis plan supporting a study of how long binding tariff classification rulings issued by US Customs and Border Protection remain in force before the agency withdraws them. The corpus comprises 221,442 rulings issued between 1989 and 2026, enumerated from the public Customs Rulings Online Search System (CROSS) as a union of eighteen search vectors continued until saturation, recovering 99.99 per cent of the rulings CROSS reports as searchable. Of these, 197,215 are classification determinations, of which 2,994 were later revoked or modified by CBP and 194,221 are right-censored at the snapshot date of 9 August 2026. Withdrawal is identified from the agency's own revocation and modification linkage fields rather than inferred, so the event definition requires no entity matching, textual similarity or model-based linkage. The deposit contains the enumerated corpus and withdrawal graph, the persistence and survival datasets, duty rates extracted from ruling text, all diagnostic outputs and console logs, the PowerShell and Python analysis code, five publication figures, and the analysis plan with its deviation log. Scope conditions. The outcome is detected error only; rulings that are wrong and never withdrawn are unobservable, so the observed withdrawal rate is a floor on error rather than an estimate of it. Coverage begins in 1989 when the Harmonized Tariff Schedule took effect, and absence of older withdrawal targets is left truncation at database inception rather than record deletion. Duty findings hold for ad valorem goods only. Timing disclosure. The analysis plan, including hypotheses, model specifications, the validation protocol, reporting bands and eight falsification conditions, was fixed before any model was fitted, but deposited after the analysis was completed. It is therefore a pre-specified analysis plan rather than a timestamped pre-registration. Corroboration is available in the code: falsification thresholds and their reporting language are hard-coded in scripts whose file timestamps predate the results they evaluate. Two conditions fired and are reported as negative results without modification to the specifications that produced them. The underlying rulings are United States Government work in the public domain. The derived data and analysis code in this deposit are released under CC BY 4.0.