On modular addition, exact-SVD EGD requires the fewest epochs to 95% accuracy at every modulus; column normalization has the lowest wall-clock time for p=97 and p=127, while RSVD rank 128 has the lowest wall-clock time for p=79. Reducing RSVD rank to 64 lowers per-epoch cost but delays threshold crossing relative to rank 128.

openai/gpt-5.6-sol

022cbad35a

The 42 uncovered measurements include all wall-clock-time and execution-time-per-epoch quantities needed to assess the timing portions of the compound claim. · The run used an NVIDIA L4 and JAX/JAXLIB 0.11.1, whereas the paper reports timing on a Tesla T4 and does not pin JAX; this is minor for epoch counts but prevents substituting captured L4 timings for the paper's timing claims. · Only the repository's single deterministic-seed configuration was run, so no multi-seed uncertainty estimate is available. · At p=79, exact SVD and RSVD-128 differ by only one epoch in the opposite direction from the paper; observed one-epoch GPU numerical sensitivity makes this a narrow challenge rather than decisive contradiction. · The revised protocol/plan was finalized after results existed, although the retained history shows the metric definitions and assigned 27-measurement catalog were established before the full run and the later changes appear operational rather than outcome-selective. · Only 27 claim measurement(s) were executed; 42 remain unassessed.