Inversion Battery v0.3 (EA-SEI-BATTERY-01 v3.0) is the ten-test successor to deposit #1455, and is as much a record of its own failures as of its findings. Each test was pre-registered with its prediction committed before its run; two predictions were confirmed, two partially confirmed, six missed, one registered panel was not run, one line of inquiry was closed with a specific handoff, and seventeen deviations between registration and execution are enumerated in the record.
Its strongest result reads three scores off a single trained encoder and finds they disagree about which class is anomalous: with model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation. That falsifies the program's own post-v0.2 claim that architecture changes the orientation of the blind spot. Alongside it: global teacher-student agreement does not guarantee preservation of the operational tail, with rank correlation reaching 0.904 on background while half the teacher's top-one-percent selections are lost and weight quantization contributes no systematic further loss; an architecture's detection behaviour is not transportable across representations of the same events, since the same classes encoded two ways reverse the ordering of the architectures entirely; complexity comparisons require a common coordinate basis, and a previously recorded reading of this data had the ordering inverted; and the constituent inversion survives hierarchical uncertainty with real calibration, its entire interval below chance.
The published normalized-autoencoder remedy was implemented on the third attempt with the construction the paper specifies, and a registered convergence diagnostic — added so that a null could be distinguished from a non-run — showed it reducing directional asymmetry on the low-dimensional pairs while failing to equilibrate on the high-dimensional constituent representation under any tested chain length or sampler geometry. That question is closed for this battery and handed on with the parameters a better-resourced group would sweep, together with the observation that sampler convergence and the inversion are correlated at 0.94, which points away from what the program predicted.
Two earlier claims are withdrawn, one arm was relabelled twice after audit rather than overwritten, four overclaims returned by Assembly review are declined with reasons, and a concurrent-writer incident that nearly overwrote the entire verdict ledger is recorded with its repair. The deposit is made in that condition deliberately: a measurement program that asks institutions to publish what their instruments fail to retain has no standing unless it publishes what its own instruments failed to do.