Experimental data: Gradient Eligibility Trace Is Momentum - A Four-Arm Controlled Experiment on Real Text

Data supporting the manuscript 'Gradient Eligibility Trace Is Momentum: A Four-Arm Controlled Experiment on Real Text'. (1) inject_v4_results_final.json: complete per-step loss curves, threshold-step records, and wall-clock timings for the four-arm experiment (vanilla SGD / SGD+trace / AdamW / AdamW+trace) x 3 seeds (42/123/999) on WikiText-2 with a 26.9M-parameter causal Transformer decoder (8,000 steps per run), including the SGD learning-rate calibration grid {0.01, 0.03, 0.1}. (2) inject_v4_final.py: the experiment script (trace decay lambda=0.9, injection beta=0.05, gradient clipping 1.0). (3) inject_20k.log: the preliminary-run log cited in the manuscript's erratum analysis, showing that the previously reported 5-20x acceleration values were ratios computed at the 1e-5 to 1e-7 loss noise floor (e.g. 13.5x at step 7,000 inverting to a 63x reversal at step 20,000). Hardware: NVIDIA V100-SXM2-32GB, PyTorch 2.4.1+cu118.