This paper is the methodological strand of the program turned outward. The mass-relation work runs on a strict internal discipline, discovery and confirmation kept separate, null models pre-registered, hit fractions quoted against explicit priors, because a single physicist pattern-matching on PDG tables is exactly the failure mode the field worries about. Machine learning industrializes that failure mode: candidate generation becomes effectively unbounded while evaluation bandwidth stays fixed.

The argument is information-theoretic rather than alarmist. Once generation outpaces evaluation, binary accept/reject degenerates to random sampling; the only response that preserves ranking under a finite budget is stratification. The paper proposes seven computationally evaluable criteria, rules that can be adjusted retroactively and thresholds that can be tuned as results accumulate, so that ML-generated candidates can be prioritized for pre-registration and longitudinal tracking rather than prematurely accepted or discarded.

The goal is infrastructure built before it is urgently needed: validation criteria for an AlphaFold-scale event in particle physics, offered as a provisional framework for community calibration.