Egalitarian Gradient Descent: A Simple Approach to Accelerated Grokking

Each cell is one deterministic source-seeded run; no replicate distribution or uncertainty interval was assigned or estimated.

Execution used an NVIDIA L4 and JAX/JAXLIB 0.6.2. The assigned measurements are hardware-invariant epoch counts and ratios; no paper wall-clock or per-epoch timing claim is inferred from this substituted timing environment.

The wider claim catalog includes timing and per-epoch quantities that were not among this run's 27 assigned measurements and remain outside execution completion.

Published Table 3 values were visible during source interpretation before execution and were not used as fresh evidence or to select outputs.

Claim coverage is partial: all 42 wall-clock, timing-speedup, and time-per-epoch measurements are uncovered, including the claims about the fastest method at each modulus and rank-128's overall cost balance.

The official main-results/mod_mult.py content was not supplied in evidenceContext, and run_mod_mult_experiment.py was truncated across the core EGD/RSVD and training-path implementation; source equivalence therefore cannot be audited.

Only one deterministic source-seeded run was executed per configuration; no replicate distribution or uncertainty interval is available.

Execution used an NVIDIA L4 with JAX/JAXLIB 0.6.2 rather than the paper's Tesla T4 environment. Epoch counts are not timing quantities, but device/library numerical differences may affect threshold trajectories; paper timing is not comparable.

The complete raw curve files were not supplied directly to the assessor. Captured jq audits establish their count, declared lengths, configurations, and threshold recomputation, but not inspection of every raw curve entry.

The plan was revised after paper values had been viewed. Its threshold and ratio definitions are source-consistent and the revision preceded the full run, but the missing implementation material prevents independently verifying the claimed faithful equivalence.

Only 27 claim measurement(s) were executed; 42 remain unassessed.

This execution covers only part of the compound claim; unassessed measurements: m-c11-79-sgd-t, m-c11-79-sgd-pe, m-c11-79-svd-t, m-c11-79-svd-ts, m-c11-79-svd-pe, m-c11-79-r128-t, m-c11-79-r128-ts, m-c11-79-r128-pe, m-c11-79-r64-t, m-c11-79-r64-ts, m-c11-79-r64-pe, m-c11-79-cn-t, m-c11-79-cn-ts, m-c11-79-cn-pe, m-c11-97-sgd-t, m-c11-97-sgd-pe, m-c11-97-svd-t, m-c11-97-svd-ts, m-c11-97-svd-pe, m-c11-97-r128-t, m-c11-97-r128-ts, m-c11-97-r128-pe, m-c11-97-r64-t, m-c11-97-r64-ts, m-c11-97-r64-pe, m-c11-97-cn-t, m-c11-97-cn-ts, m-c11-97-cn-pe, m-c11-127-sgd-t, m-c11-127-sgd-pe, m-c11-127-svd-t, m-c11-127-svd-ts, m-c11-127-svd-pe, m-c11-127-r128-t, m-c11-127-r128-ts, m-c11-127-r128-pe, m-c11-127-r64-t, m-c11-127-r64-ts, m-c11-127-r64-pe, m-c11-127-cn-t, m-c11-127-cn-ts, m-c11-127-cn-pe.

Only 27 claim measurement(s) were executed; 42 remain unassessed. Runner records demonstrate a current execution of all 15 declared 15,000-epoch configurations and correct extraction of the 27 epoch and ratio fields. Under the executed wrapper, exact-SVD was earliest at p=97 and p=127, but RSVD-128 was earlier at p=79 (222 versus 228 epochs). However, the official source file and the omitted middle of the adapted runner—including the complete EGD/RSVD update and training path—were not supplied for audit, so equivalence to the paper computation cannot be verified. Moreover, 42 timing and per-epoch measurements were not executed, leaving the wall-clock and overall-balance portions of the claim uncovered. Numeric proximity therefore cannot establish reproduction of the full claim.

Claim coverage is partial: 42 timing and per-epoch measurements were not executed, so the wall-clock winners and rank-128 cost/step-count balance are unverified.

The rerun used an NVIDIA L4 with JAX/JAXLIB 0.6.2, whereas the paper reports timing on a Tesla T4. This does not invalidate epoch comparisons, but it can alter floating-point trajectories and cannot support the paper's timing claims.

Each cell is a single source-seeded run; no replicate distribution or uncertainty interval is available to determine how stable the six-epoch p=79 ordering reversal is.

The plan was revised before the full workload after source review: the non-executable, hard-coded combined script was replaced with a faithful single-configuration wrapper and deterministic aggregator. The revision exposed configurations and serialization without intentionally changing the update equations.

The retained corpus does not expose the 15 large per-epoch JSON files for direct inspection. Their existence and internal checks are evidenced through the complete run log and captured validation commands; this is platform-level evidence loss rather than an identified author omission.