Independent reproduction · arXiv 2605.01082v1 · OpenReview mrtg4NmvAe
Networked Information Aggregation for Binary Classification
Exact BCE identities, sequential logit-passing, and depth-bottleneck evidence with primary sources pinned.
22/22
scientific gates
48
network panels
56
lower-bound cells
48/48
probability control worse
Scope: The upper result requires M-coverage and bounded features/coefficients. The lower construction has D=kp and p≤k−1; “matching” means the depth bottleneck, not identical exponents. The protocol passes logits, not probabilities.