experiment: opening-pause trim × photometric augmentation (12 ACT runs)

[parvpatodia]

Adds Parv's experiment record to experiments/, plus the standalone audit script it depends on.

Sibling to 2026-08-04_act-8bin-6run. This is the top+front arm — the only one of the three owners' grids without the physical wrist camera.

Two axes the 6-run grid left out

1. Opening-pause trim. The 6-run eval protocol predicted "expect a hesitant start at block-A positions… the fix is trimming dead time at collection." This tests it without re-collecting.

• 8,722 of 72,518 frames (12.0%) are pre-teleop dead air. Left episodes average 3.53 s, right 1.38 s.
• A 20× change in the departure threshold (0.5° → 10°) moves the cut by 1.5 percentage points — evidence it lands on genuine dead air, not slow task motion.
• Interior pauses (4 of 119 episodes) and trailing pauses (117 of 119) are measured and deliberately left alone.
• Nothing on disk or on the Hub is modified. The trim is a SubsetRandomSampler index list, published as Parv-09/phi_so101_8bin_v1_trim.

2. Photometric augmentation. Helps weakly and consistently (gap narrowed 6/6, ratio 6/6, val 4/6) but at matched steps raises train L1 only 0.5–6%, so it is too weak to really regularise — for a 49% throughput cost.

First robot result

✅ The arm no longer freezes at start. ❌ It still cannot localise the duck. Per-bin success rates pending; the untrimmed control has not been run yet.

Independently reproduces the 6-run finding

Posterior collapse: raw KL 0.0969 at epoch 1 (81% of total loss) → 3e-5 by epoch 31, against the 6-run probe's kld_loss = 6.7e-5. Two arms, same conclusion.

⚠️ Reviewer notes

• Losses in this file are NOT comparable with the 6-run file. This grid carves a 12-episode in-distribution val set out of the training pool (77 train, not 89). The held-out 30 are identical and untouched in both.
• Everything is n=1 seed. No claim here should go in a paper or the Research Day poster as-is.
• The file also records four corrections to earlier claims I made in this project, and an scp truncation trap that left a 163 MB stub of a 206 MB checkpoint while exiting 0.

Scope

src/phi/utils/deadtime.py is included because the experiment links it and it is standalone (verified running, with a guard against partially-fetched caches). The trainer and gap_eval.py are not here — the eval script imports split and trim helpers from the trainer, which needs packaging work. Follow-up PR.