exp: E28 - under DPPO, dppo-policy learns HalfCheetah and climbs Hopper/Walker2d; fpo-policy does not move

[tactino]

E24 (#59) ran every DPPO combination on MuJoCo for twenty iterations and claimed only that each runs; unregistered, dppo-policy rose on Hopper and Walker2d while fpo-policy under the same DPPO did not move. E28 runs the same five cells, through E24's own run_cell.sh unchanged, to 100 iterations (409,600 steps), with a rule for "learns" fixed in advance. Stacked on #59.

Result

cell	iterations 91-100, seeds 0 / 1 / 2	status
dppo-policy · DPPO · HalfCheetah	+481 / +84 / +544 over the first	learns (2 of 3 past +200)
dppo-policy · DPPO · Hopper	339 / 430 / 276 (from 11-16)	did not learn in 409,600 steps (bar: 500)
dppo-policy · DPPO · Walker2d	260 / 292 / 274 (from about −2)	did not learn in 409,600 steps
fpo-policy · DPPO · Hopper	25 / 25 / 27 (from 15-20)	did not learn in 409,600 steps
fpo-policy · DPPO · Walker2d	3 / 1 / 2 (from 0-2)	did not learn in 409,600 steps

• P1 holds, 5 of 5 (15 seeds, 100 iterations each, no traceback, client exit 0).
• P2 and P3 falsified: I extrapolated E24's twenty-iteration slope. Walker2d flattened after about fifty iterations; Hopper was still climbing (best iteration 493).
• P4 holds: fpo-policy under DPPO moves by +1 to +11.
• dppo-policy learns HalfCheetah - the one cell registered without a prediction, flat for its first twenty iterations.

The registered joint reading ("the policy class decides") needed P2 or P3 and is not claimed. Described instead: with the algorithm, variant, data per iteration and optimiser steps per iteration all equal, dppo-policy gains +84 to +544 across the three tasks and fpo-policy +1 to +20. fpo-policy learns all three under FPO, so by the project's rule its standstill under DPPO is our defect to locate. One code-read candidate, untested: dppo-policy floors every denoising step's noise at 0.1, while fpo-policy's flow steps run from 0.1 down to about 0.01 (E18's candidate, now with the code's numbers).

Every seed's first iteration matches E24's to the decimal: collection is deterministic, learning is not (E20).

Matrix after E24, E27 (#62), E28

	HalfCheetah	Hopper	Walker2d	robomimic square
fpo-policy · FPO	learns	learns	learns	runs
fpo-policy · DPPO	runs, no learning	runs, no learning	runs, no learning	runs
dppo-policy · DPPO	learns	runs, rising: 276-430	runs, rising: 260-292	runs

Files

• PROTOCOL.md (pre-registered, d1f4783, before any run), FINDINGS.md
• run.sh (calls E24's run_cell.sh), summarise.py
• summary.tsv - one row per cell and seed, in E24's columns
• results/ - every server and client log, verdicts.txt; checkpoints and tensorboard stay on the workstation