Pre-registration DRAFT: the ch0 isolation cell (demos + shoulder-pan-affined clean)

Draft cut 2026-08-22 06:5xZ (work session), the gripfix pre-reg’s registered ≤10-branch follow-up — carrier-hunt ladder rung 2, drafted at the gripfix verdict (5/100 — gripper amplitude NOT the sole carrier). The registered follow-up named a ch0 shift; the constant-freeze read below (spec frozen in-channel 06:45Z/06:47Z BEFORE compute, posts 1540612893836574780 / 1540613448264712243) refuted the shift form at zero GPU cost and certified a moment-matched affine as the only viable one-channel edit — this draft registers that form amendment with the measured evidence, before any launch. Launch is delegated per the standing no-GO-ask rule: the cell fires at a free GPU window once the materializer’s oracles are green, announced in-channel at launch. DRAFT status: the materializer named below does not exist yet; the reads and grid below are frozen now and do not move.

Plain words. Seven episodes of real-robot data poison grasping at 0.7% of the training mix, and we are hunting the carrier one suspect at a time. The loudest suspect — a gripper that never opens all the way — was fixed and grasping stayed collapsed, so it isn’t (only) that. The second-loudest suspect is the shoulder-pan joint: those seven episodes barely swing it. We first checked whether the joint is simply offset (pointing somewhere slightly different on average) — it isn’t: the average position matches everyone else’s to within 1.4 units. What differs is the range: the episodes sweep about a third as widely as every other dataset. So this experiment stretches that one joint’s recorded values — same center, 2.76× the swing — so its statistical footprint matches the rest of the data exactly, changes nothing else, and retrains. If grasping recovers, the shoulder-pan distribution was the poison. If it stays collapsed, both named suspects are exhausted and we move to testing the seven episodes one at a time.

The measured basis (banked this session, CPU, record-only)

Instrument fontaine/scripts/ch0_shift_constant_read.py + ch0_affine_addendum_read.py, report reports/analysis__ch0_shift_constant_read.json; oracles green (banked manifold-probe KS reproduced to 1e-9; KS shift-invariance under equal shift).

• The “modest ch0 shift” suspect was mis-specified. Clean’s ch0 mean sits 1.39 raw units from demos’ (+1.48 vs +0.09 — ≈0.05 of demos’ std). The anomaly is spread compression: std 10.16 vs demos 27.99 and v2 19.21; q05/q95 span ±14/17 where demos sweep ±46.
• A pure location shift is pre-refuted: both frozen candidates (Δ_mean −1.39, Δ_median +1.39) leave KS essentially unmoved — 0.295 → 0.286 / 0.308 vs the demos↔v2 reference 0.161. Launching a shift cell would have burned ~17 GPU-h on a transform measurably unable to move the marginal.
• The moment-matched affine lands deep in-band (frozen decision rule from the addendum spec): post-affine KS vs demos 0.047 action / 0.049 state (reference band 0.161/0.159); vs v2 0.193/0.206 (slightly over the band on the v2 side — recorded, the rule keyed on the clean↔demos side). The robust (median/range95-matched) candidate also lands in-band (0.092/0.094) but strictly worse; moment wins.
• Support containment: the transformed range [−72.9, +50.6] sits inside demos’ observed ch0 support [−110.0, +79.6] (action and state) — the edit does not manufacture out-of-support joint values.

The cell

Mix = grasp_demos_v2/merged + so101_pick_place_clean_ch0fix_n (×4), and nothing else — where clean_ch0fix_n is the 7 clean episodes with the shoulder-pan channel (ch0, action AND state) transformed by the frozen affine

x′ = 0.0923439813196304 + (x − 1.481974338423806) × 2.755193138766973

(target = demos’ action-column moments, the shared-convention choice gripfix made with 41.69; one transform for both columns, constants from the action stats, state-vs-action deltas banked for reference). All other channels byte-identical; episode count, lengths, camera streams, annotations untouched.

Frozen transform decisions (no exec degrees of freedom): the moment-matched affine above, exactly — not per-episode, not re-estimated; ch0 only; both action and state columns (state-action consistency for the state-conditioned policy, the same rule gripfix froze); constants pinned verbatim to the banked read. --recompute-stats then gives the transformed set its own pdnorm row.

Dataset name frozen: mcobzarenco/so101_pick_place_clean_ch0fix_n — the suffix is chosen (gripfix Amendment-1 lesson, pre-applied) so the episode holdout draw is exactly (2,): holdout_episodes(repo_id, 7, 0.1, 0) verified this session — the clean-side train split stays episode-identical to democlean’s ({0,1,3,4,5,6}, 3026 kept frames, 0.69% share, the same 373-frame episode held out).

Materializer: a new fontaine/scripts/make_clean_ch0fix_dataset.py (exec session) with oracles: ch0 columns equal the affine of source exactly (float64 transform, cast back to source dtype); every non-ch0 column byte-identical; frame/episode counts identical; transformed range inside demos’ observed support [−110.0, +79.6]; recomputed holdout draw is (2,); a no-op guard that refuses if the source ch0 std already exceeds 20 (i.e. the source is not spread-compressed).

Command

The democlean launcher verbatim (launch_local_grasp_sft_v2_joint_1gpu_pdnorm_democlean_h100.sh) with exactly one delta: so101_pick_place_clean → so101_pick_place_clean_ch0fix_n in --train-data (the --dataset-repeat 'mcobzarenco/so101_pick_place*=4' glob matches unchanged). Per-dataset flow norm, joint --insulate-flow ce-weight 1.0, --recompute-stats, eff-batch 96, decoder-lr 5e-5 / backbone-text-lr 1e-5, image-augment 0.8, holdout 0.1, eval-250, save-500 + keep-latest-optimizer pruner from launch, 3000 steps, seed 0 (same-seed comparability with all five anchors). Fit smoke first; compute-app abort guard carried.

Baselines and anchors (all banked)

• democlean 8/100 — the direct paired arm: same 7 episodes, un-transformed. THE comparison.
• gripfix 5/100 — the exonerated-suspect anchor: same discipline, ch5 edit, no recovery.
• Onerig 28/100 — the healthy two-dataset cell.
• Demosonly control 11/100 — the never-mixed floor.
• Convicted three-way 1/100 — the original collapse.

Reads and frozen decision grid

Primary — sim100 flow leg at step 3000, protocol byte-identical to the democlean/onerig/convicted/gripfix batteries (unseen seeds 0–99, euler-10, episode 30 s, execute-horizon 30, bfloat16 decoder, --stats-repo-id grasp_demos_v2/merged, --clutter-appearance standins).

Same numeric grid, interpretations adapted:

• ≥ 20/100 → the ch0 distribution IS the carrier: a one-channel affine de-poisons the 7 episodes. Banks same-session; the mix-hygiene implication extends from amplitude conventions to per-channel spread checks; clean re-enters the pool transformed.
• ≤ 10/100 → the ch0 marginal is NOT the (sole) carrier, and the named-suspect list from the manifold probe is exhausted. Named follow-up (delegation standing): per-episode leave-one-out at ×4 — 7 cells, the ladder’s registered last rung — with a design pass first (the LOO pre-reg should weigh cheaper orderings, e.g. leave-K-out bisection, before committing ~17 GPU-h × 7).
• 11–19 → ambiguous (the control’s band): per-channel MAE, per-slice breakdown, videos to the owner, no claim.

Registered asymmetry note (honesty clause, priced in before launch): this cell is one-sided. ≥20 is decisive — an edit that recovers grasping proves the carrier. ≤10 is weaker than gripfix’s ≤10 was: the ×2.755 amplification enlarges the vision–proprioception disagreement class that gripfix’s ≤10 branch flagged (its remap certifiably hurt paired progress, −2.07 cm CI-excl-0), so a no-recovery outcome cannot fully distinguish “carrier remains” from “edit-added artifact re-poisons”. The grid verdict stands as written (band membership), but the ≤10 mechanism claim is capped at “suspect list exhausted, edit-family exonerated as a fix path”. The paired-progress read carries the artifact detection either way.

Paired reads recorded alongside (sim100_paired_read.py): vs democlean 8/100 (THE read), vs gripfix 5/100 (edit-artifact comparison: two one-channel edits, same discipline), vs onerig 28/100 (yardstick), vs control 11/100.

Secondary — drift guard: Δeval(1000−500) ≤ +0.30, record-only (the probe canNOT clear this cell — decoupling banked three times).

Tertiary — panel guard at endpoint vs the disc-1000 banked npz, frozen house rule (fail = worse than +0.05 CI-excl-0), truthfit rewear alongside — unchanged from the last four cells.

Record-only: the recomputed clean_ch0fix_n pdnorm row vs clean’s banked row (ch0 scale should move ×2.755, everything else pinned — the live materializer oracle); eval-250 curve vs the democlean curve (same-seed, one-channel-delta twins).

Gates and boundaries

• GPU-hours gate: 17 (train ~13.7 measured on the identical recipe + battery ~3).
• Launch window: delegated, announced in-channel, never gated on an owner GO. The squint screen owns the H100 first (leg B → ~10:3xZ 08-22, then leg C ≈1 GPU-h); this cell fires at the first free window after the squint leg C boundary unless owner steering reorders.
• Boundaries: step-1000 drift read (record-only); step-3000 endpoint → the battery script pattern (clone launch_gripfix_endpoint_battery.sh, name deltas only) → verdict through the grid.
• Checkpoint policy: saves under ~/checkpoints/finetune/grasp_sft_v2_joint_1gpu_pdnorm_ch0fix; endpoint banks weights-only on any decisive read.
• Seed policy: seed 0 (comparability); ×4 repeat and the 0.69% effective share carried unchanged.

Objection/decision path: reads and grid frozen as drafted; the exec session’s only freedoms are the materializer’s file-format details and the launch window. Any spec change lands as a registered amendment first.