ProCreations/repro-shift-dependent-asymmetry-inverse-lora-federated-medical-segmentation

Reproduction bundle

Immutable deployment-slug notice: this package is deployed into a pre-existing reserved shell whose immutable slug (repro-shift-dependent-asymmetry-inverse-lora-federated-medical-segmentation) does not describe this paper. The canonical identity is OpenReview nDfDnsyllY, arXiv 2603.00023v1, Riemannian Dueling Optimization; the unrelated slug is capacity infrastructure only.

Pinned arXiv v1 PDF and complete source archive, the mandatory semantic-v4 evidence matrix, 44 deterministic method gates plus 9 literal VGG/CIFAR gates, two warning-strict byte-identical runs, and claim-specific destructive controls for all six registered claims.

PYTHONDONTWRITEBYTECODE=1 PYTHONHASHSEED=0 PYTHONWARNINGS=error OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 python3 -W error reproduce.py --out /tmp/rdo-replay
PYTHONDONTWRITEBYTECODE=1 PYTHONHASHSEED=0 PYTHONWARNINGS=error python3 -W error reproduce_vgg_cifar.py --device mps --out /tmp/rdo-vgg-replay
PYTHONDONTWRITEBYTECODE=1 python3 validate_evidence.py
PYTHONDONTWRITEBYTECODE=1 python3 verify_manifest.py

The executed suite includes 132 direct increasing-scale RDNGD rate runs; 60 proof-scheduled recurrent runs over four dimensions; 32 increasing-scale growing-batch RDFW runs; 840,000 random tangent directions; 81 SO(2) leveling problems; a literal released 28.1M-parameter VGG11-BN on actual CIFAR-10 bytes with a 3,072-dimensional, 1,000-step, batch-10 comparison-only attack; the paper's 50,000-step Rayleigh runs at d=100 and d=150; and 5,000-step Karcher runs at n=5 and n=10 with 50 SPD inputs and condition-number span 10^6.

Important boundary: the displayed RRDNGD theorem writes epsilonk with a `2^(2-k)` denominator, which grows with k. Algorithm 2 and the proof instead set Dk=D/2^k and epsilonk=alpha Dk/4=alpha D/2^(k+2). The executable uses the proof-consistent schedule and records this source inconsistency rather than silently widening the theorem.

The VGG/CIFAR execution uses the actual released model state dict and dataset row named by the paper-linked reference code. Its successful eta0=0.8 cosine schedule is a labeled clean-room choice, not an author setting. The paper's displayed eta=1e-6 is executed as a boundary control and does not flip the class.