E43: training with the env clients on another physical machine, and what the crossing costs

[tactino]

E43 trains the coverage figure's fpo-policy · FPO · HalfCheetah cell with its env clients on a second physical machine, and measures what that crossing costs. The server ran on guangzhao (wired Linux). The env clients ran natively on a Windows laptop on campus Wi-Fi, over Tailscale peer-to-peer. PROTOCOL.md was pre-registered at 46bce03 after the pilot. AMENDMENT.md (cec6675) was written before its runs.

Every registered prediction holds (results/verdicts.txt):

• Training (P1-P3). Both arms run end to end, and all 3 seeds learn in each. Over iterations 91-100 the gain over iteration 1 is +1,555 / +1,020 / +976 locally and +2,404 / +2,445 / +1,196 across machines, against a bar of +200.
• Wall clock (P4). Across machines each run took 1,253-1,356 s longer, against a pilot-based 1,119 s ±25%. All three runs sit above the point estimate.
• Ladder (P5, P6).
  • On one machine, loopback and the machine's own Tailscale address agree.
  • Between the machines, cost = a fixed ~2.8 ms + 2 x the observation's bytes / throughput (18 MB/s). Measured values come in 3-10% under the predicted ones at 48, 184 and 588 KiB.

The cross arm ended higher on 2 of 3 seeds, which nothing predicted. The amendment added seeds 3-5 to both arms. Six seeds per arm read as seed variation: the means are 371 apart, against an sd of 694 within an arm. The figure shows each arm's band of seeds and no mean line.

It corrects E10. "1.3-3.6% of a step" used E7's VM-to-host crossing. Over this link a 184 KiB observation costs 21.4 ms per exchange, 21% of pi0.5's 100 ms forward. The README's evidence table and experiments/README.md's missing-items entry now say so. E7's row there is replaced by E43.

The laptop's MuJoCo differs from guangzhao's in the last bit from the first step (8.9e-16), so the arms are compared by the status rule, not trajectory by trajectory. On one machine a seed reproduces exactly.

Files: experiments/e43-cross-machine-training/, README.md (the evidence table), experiments/README.md (the missing-items entry).