FlowWM/causal_droidstd/base_vs_causal/README.md

# BASE Ctrl-World vs CAUSAL checkpoint-20000 — same 4 val episodes, same seed

`SIDEBYSIDE_*.mp4`: TOP half = base, BOTTOM half = causal. Within each half: GT over prediction.

| val | task | BASE | CAUSAL-20k | delta |
|---|---|---|---|---|
| 99  | pull the wooden handle | **19.28** | 16.92 | -2.36 |
| 199 | move the green towel   | **20.68** | 16.86 | -3.82 |
| 299 | (unlabelled)           | **23.05** | 18.42 | -4.63 |
| 399 | wipe with a napkin     | **16.92** | 15.22 | -1.70 |
| | **mean** | **20.0** | **16.9** | **-3.1** |

Base wins on 4/4, by 1.7-4.6 dB. That is a systematic per-episode gap, not the 3.2 dB
episode-to-episode spread, so it is a real effect.

## Two competing explanations -- NOT separable from this data

(a) The Phi branch is costing appearance quality (the original concern: a shared UNet asked
    to emit both a sharp frame and a smooth flow field).
(b) Plain fine-tuning loss: base was trained on the FULL DROID corpus; the causal run
    fine-tunes on 387 trajectories, ~200x narrower. Any fine-tune on a set that much
    smaller degrades PSNR even with no Phi at all (catastrophic forgetting).

Deciding between them needs the ablation-table `baseline` row that has never been run:
Ctrl-World fine-tuned on these same 387 trajectories with lambda_flow=0. Until that exists,
this table does not show that the causal design is harmful -- only that this checkpoint is
worse than the corpus-trained base.

Also note the causal run is at 33% of its schedule (20k/60k).