For every sparse-parity k, at least one RSVD configuration reaches 95% accuracy in fewer epochs and less wall-clock time than exact-SVD EGD. The winning table rows are RSVD rank 40 for k=2 and rank 30 for k=3 and k=4. Column normalization and RSVD rank 40 do not reach 95% for k=4.

openai/gpt-5.6-sol

ace3a41be6

Claim coverage is partial: only 35 of 91 paper-side measurements were assigned, leaving 56 uncovered; the overall claim must therefore remain inconclusive. · The paper describes hinge loss, while the official released function and rerun compute mean squared error plus L2. This is a scientifically material protocol difference. · No comparable Tesla-T4 time-to-95 measurement was executed. Full-horizon elapsed times on an NVIDIA L4 do not establish the claimed wall-clock ordering. · The same epoch advantage could occur while RSVD is slower in wall-clock time because RSVD and exact SVD have different per-epoch costs; the missing time-to-threshold comparison is therefore discriminating evidence, not paperwork. · Under the released squared-loss computation, k=4 rank 40 did not cross, but column normalization crossed at epoch 4500, opposing the claimed column-normalization failure without conclusively contradicting the hinge-loss paper experiment. · Each condition used one fixed source-seeded trajectory, with no across-seed variability estimate. · The post-result aggregation revision allowing null values for genuine non-crossings is supported by the metric definition and was not an outcome-selected replacement value. The qualitative k=4 controls remained in scope. · The 21 referenced per-epoch raw run JSON files were not retained in the accessible evidence corpus. Runner stdout, aggregate data, and the executed validator summary remain, but independent point-by-point inspection of those histories was not possible. · Only 35 claim measurement(s) were executed; 56 remain unassessed.