Inversion Battery v0.2 (EA-SEI-BATTERY-01 v2.0) is the multi-seed successor to deposit #1449, re-running the pre-registered inversion battery with five independent seeds per cell across sixteen cells and eighty trained model pairs, on the public top-quark tagging reference dataset and the LHC Olympics 2020 R&D dataset.
Its surviving findings: the top/QCD inversion reproduces across five trainings — a QCD-trained autoencoder detects top jets at AUC 0.838 while the reversed system rates QCD as more ordinary than its own training class at 0.243, and the density family inverts likewise (0.873 / 0.319), so the failure is not a property of reconstruction loss; direction-dependence appears across all tested families on the background/signal pairs with seed-stable intervals; the near-structure control stays an order of magnitude quieter; and at the severe operating point every system assimilates most of the partner class, so absolute retention and directional asymmetry are distinct quantities. The encoder-side latent score alone escapes the inversion while carrying the largest asymmetry on that pair — architecture changes the orientation of the blind spot. From the distillation test: global student-teacher rank agreement is moderate while only 36 to 47 percent of the teacher's top one percent survives into the student's top one percent, and weight quantization costs almost nothing further.
The deposit is as notable for what it refuses as for what it reports. A seven-item correction ledger records every mismatch between registration and execution — an unfaithful normalized-autoencoder surrogate that leaves the registered remedy prediction untested, a per-class rather than pooled complexity measurement that leaves the Low-Complexity Blind-Spot Hypothesis inconclusive, distillation measured on the withheld class rather than background, a weight-only quantization surrogate, a pilot-scale miss-overlap measurement, seed-level rather than hierarchical intervals, and the finding that the inversion asymmetry is uninterpretable without the mean assimilation burden beside it. Four readings returned by Assembly review are recorded as rejected, including that the run proves the thesis, that the complexity hypothesis is confirmed, and that the normalized-autoencoder remedy fails.