An earlier step of the lowcomm-training programme found that 4 workers averaging every 16 steps (H=16) reached a slightly lower validation loss than synchronous AdamW on a small character-level GPT. A fair worry was that the synchronous baseline sat between grid points.
We added learning rates 1.4e-3 and 2.8e-3 for both methods (20 new runs, two tuning seeds, three report seeds). Both methods still pick 2e-3, and both get worse on either side, so the baseline was not under-tuned. At each method's own pick, H=16 is 0.0139 nats better on average (1.5758 against 1.5897, better on all three seeds).
Limits: one small model and corpus, plain averaging, no outer momentum. Log and code: research/lowcomm-training/LOG.md, milestone 3f.