Assimilation Across Accelerator Classifier Architectures (EA-SEI-ACRB-01 v1.0) defines the Accelerator Classifier Retention Battery, the cross-family measurement framework of the accelerator selection-metrology program, deposited under the Nobel Glas heteronym.
ACRB is a pre-registered framework for determining whether the directional assimilation measured in reconstruction autoencoders persists across materially different learned-selection architectures: encoder-side latent scores, explicit density estimators and normalizing flows, normalized autoencoders and WNAE, distilled anomaly triggers, and supervised DNN and GNN trigger classifiers entering through open-set assimilation tests. It compares architectures at matched own-background operating points, follows models through deployment transformations — teacher-to-student blind-spot inheritance, float-to-firmware retention drift — and measures the topology and overlap of miss regions rather than discrimination alone.
The design is symmetric in outcome by construction: if symmetric-by-design architectures remove directional asymmetry across the pre-registered pair set, ACRB is the validation instrument for that remedy; if assimilation persists across families, the evidentiary burden shifts from a single-model defect toward a selection-system phenomenon. Its claim boundaries are explicit — assimilation is operational, AUC and score-order inversion and threshold-level retention are distinct estimands, public benchmarks do not establish deployed experiment efficiencies, and architectural diversity does not imply diversity of failure. Cross-model miss overlap is measured as the Representational Independence Index family of #1450.