Topology-aware machine-learning software and a preregistered graph-classification research program spanning 14 hypotheses (H001-H011b) and 53 falsifiable sub-predictions. Under the tested matched-capacity configurations, no unique L_0 Hodge advantage is supported once the external-residual architecture is matched. A narrow NCI1 positive difference over the matched MLP remains; the H009-R corrective replication did not detect a learned-sheaf improvement over fixed Hodge at the tested power, and H011b remains unresolved. The repository includes persistent-homology features, differentiable topology losses, simplicial and Hodge operators, a CI-gated test suite with 100% line and branch coverage on the library package, BCa bootstrap intervals, paired Wilcoxon tests with Benjamini-Hochberg correction, and explicit retention of null, invalidated, and inconclusive results.
Santiago Maniches· Zenodo (CERN European Organi...· 0 citations
Topology-aware machine-learning software and a preregistered graph-classification research program spanning 14 hypotheses (H001-H011b) and 53 falsifiable sub-predictions. Under the tested matched-capacity configurations, no unique L_0 Hodge advantage is supported once the external-residual architecture is matched. A narrow NCI1 positive difference over the matched MLP remains; the H009-R corrective replication did not detect a learned-sheaf improvement over fixed Hodge at the tested power, and H011b remains unresolved. The repository includes persistent-homology features, differentiable topology losses, simplicial and Hodge operators, a CI-gated test suite with 100% line and branch coverage on the library package, BCa bootstrap intervals, paired Wilcoxon tests with Benjamini-Hochberg correction, and explicit retention of null, invalidated, and inconclusive results.
Santiago Maniches· Zenodo (CERN European Organi...· 0 citations
