Apppublic

ProCreations/repro-provably-data-driven-lagrangian-relaxation-for-mixed-integer-linear-programming

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes

App README

Reproduction bundle

Deterministic Trackio audit of all six exact registered claims for OpenReview RwiGcN2feP. The registered numbering and empirical references are pinned to immutable arXiv 2404.06106v1; current v3 is separately pinned and revision-mapped.

bash

PYTHONWARNINGS=error python3 reproduce.py --output replay
PYTHONWARNINGS=error python3 validate_evidence.py
python3 build_manifest.py --verify

The bundle executes the exact Hessian, Papyan-decomposition, gradient, and weight objects with destructive boundary controls. It also includes paired byte-identical paper-scale runs for:

—the Deep Linear UFM (K=3,d=60,n=40,L=5,l=3, one million epochs), and
—the ReLU Deep UFM (K=3,d=65,n=40,L=5,l=4, one million epochs).

An independent NumPy oracle reconstructs the ReLU layer-W4 Jacobian, exact nonzero Hessian spectrum, and gradient projection. The earlier reduced K=3,d=14,n=16,L=4 proxy is retained only as history and is not counted. Claim 5's unsupported “about 0.2” initial alignment remains an explicit literal falsification rather than being silently rounded into a match.