Exact theorem checks
561
C1 components · formal all-posterior proof
ICML 2026 reproduction audit · gFPPTokv9C
Expectation Consistency Loss
A six-claim source-pinned audit: two general formula results verify, the hard-bin rate reproduces, the published mini-batch gradient theorem is falsified, and unreleased benchmark evidence remains explicitly open.
Rate experiment
720
ECL slope −0.4881 · ECE −0.5102
Verified suite
109
Deterministic tests · local CPU · USD 0
| Claim | Finding | Decisive boundary |
|---|---|---|
| C1 · EC theorem | Verified | Common score support; source calibration is needed for absolute target calibration |
| C2 · Sample complexity | Hard-bin rate supported | Appendix G proof incomplete; soft ratio remains open |
| C3 · Unbiased gradient | Falsified as stated | Eq. 8 gradient 0 vs profiled Eq. 10 1/4; four independent proof failures |
| C4 · Three paradigms | Formula-level verified | Canonical/class-wise/top-label semantics executed separately |
| C5 · Digit Table 2 | Source-only | No raw ten-run outputs, checkpoints, or faithful released digit pipeline |
| C6 · Broad evidence | Partial | Capability surface pinned; PACS/ImageNet pipelines and raw results absent |