Baseline Capture Architecture (EA-SEI-BCA-01 v1.0) is the constructive specification of the accelerator selection-metrology program, deposited under the Nobel Glas heteronym.
It proposes a statistically interpretable control plane for instruments performing irreversible learned selection before durable storage, separating four functions usually conflated: a content-independent probability sample taken before the audited selection; a full-rate tap of the exact representation presented to the selector; probability-weighted enrichment channels that raise coverage of rare regions without sacrificing inferential validity; and shadow selectors whose decisions are recorded but do not control acquisition. Together they constitute a replay bank supporting retrospective estimation of retention surfaces, directional assimilation, model-to-model miss correlation, and selection drift across classifier generations.
Its central principle is deliberately minimal: under a finite storage budget, no content-sensitive algorithm can be a less assumption-laden baseline than a probability sample whose inclusion mechanism is independent of event content and whose inclusion probability is known. Sketches, random projections, and alternative neural models are argued not to qualify. The paper's sharpest formulation is that it does not propose an alternative trigger but the missing control group for a trigger — discovery selection separated from selection metrology.
Feasibility rests on existing CMS infrastructure: Zero Bias data used to train AXOL1TL, Level-1 Data Scouting capturing trigger information at the full 40 MHz while bypassing ordinary Level-1 selection, and the Global Trigger test crate receiving live inputs without determining readout. The contribution is to bind such practices into one architecture whose purpose is measuring what irreversible learned selection fails to retain.