ProCreations/repro-why-deep-jacobian-spectra-separate-depth-induced-scaling-and-singular-vector-alignment

Reproduction bundle

Deterministic CPU-only audit of all six exact registered claims for OpenReview 2kSBDoP1rE and arXiv 2602.12384v2.

PYTHONWARNINGS=error python3 reproduce.py --output-dir replay --mnist-root ../../data
python3 build_manifest.py --verify

The bundle evaluates closed-form Lyapunov exponents, finite-depth terms, separation-induced alignment, balanced and unbalanced singular-value ODEs, the exact depth-10/width-64 rank-10 synthetic setup, and two complete 60,000-example MNIST runs at width 128. The advertised anonymous author-code endpoint returned HTTP 401, so the implementation is independent and that scope is explicit. Registered Claim 5 is honestly falsified as literally worded: the source compares Lyapunov predictions at initialization, not a training trajectory converging toward them.