Basin-Insensitive Reliability: When Riddled Geometry Makes Reproducibility Unattainable, Certification Is the Correct Instrument

A full preprint synthesizing riddled-basin geometry (Ly & Gong, Nature Machine Intelligence 2026) with PAC-Bayes certification: when training outcomes are unpredictable at the seed level, the correct reliability instrument is a bound that holds for the trained model itself. Measured: a three-round governed distillation arc (0.3%, 2.7%, 1.3% gate-assisted - assistance made acceptance worse), a null uncertainty-exponent measurement with its design flaw named, and the synthesis that certification is basin-independent by construction.