Pre-registration DRAFT: carrier-hunt rung 3, both contingent branches (drafted ahead of the ch0fix verdict)

Draft cut 2026-08-22 23:3xZ (work session), riding the live pdnorm_ch0fix train (rung 2, done ~11:2x–11:4xZ 08-23) so the verdict session executes instead of drafting — queue item carrier-hunt-rung3-prereg, the ch0-affine pre-reg’s two registered branches, both drafted BEFORE the verdict exists. Exactly one branch fires, selected mechanically by rung 2’s frozen grid: branch A if ch0fix lands ≥20/100, branch B if ≤10/100; the 11–19 band fires neither (per-channel MAE + slices + videos to the owner, no claim, no launch). Launch is delegated per the standing no-GO-ask rule: the selected cell fires at the verdict session once its materializer’s oracles are green, announced in-channel. DRAFT status: neither materializer exists yet; the reads, grids, split, and dataset names below are frozen now and do not move.

Plain words. A training run is going right now that will tell us whether stretching one joint’s recorded values fixes seven poisonous robot episodes. This page writes down — before that answer arrives — exactly what we do in either case, so tomorrow’s session can act instead of deliberate. If the stretch works (branch A): the fix touched two copies of the joint’s values — the commanded positions and the measured ones — and we ask whether fixing only the commanded half suffices, because a fix at the output side alone is something any data pipeline could apply mechanically. If the stretch fails (branch B): every named suspect channel is then exhausted, and we start splitting the seven episodes themselves — train on half, see if the poison came along. We measured tonight that the joint anomaly is spread evenly across all seven episodes, so the split is chosen by frame balance, not by suspicion, and it halves the search space for the same price every earlier experiment paid (~17 GPU-hours). Either way, nothing launches until the current run’s verdict picks the branch.

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

Instrument fontaine/scripts/carrier_rung3_basis_read.py, report reports/analysis__carrier_rung3_basis.json; sanity oracles green (demos ch0 mean/std reproduce the rung-2 affine constants to 1e-12; the holdout search reproduces democlean’s and clean_ch0fix_n’s verified (2,) draws).

The ch0 anomaly is episode-uniform, not episode-concentrated. Per-episode ch0 action spread and KS vs demos, all 7 episodes:

Every episode’s std sits at 5.7–11.4 vs demos’ 28.0 — the compression is a property of the whole collection, not of an outlier episode (the gripper shortfall likewise: every ch5 max ≤ 32.3 vs the 41.69 convention). Two consequences frozen into the branches: (a) if ch0 IS the carrier, per-episode slicing could never have found it — the bisection ladder is correctly sequenced behind the channel edits; (b) if ch0 is NOT the carrier, no measured channel ranks the episodes, so branch B’s split is frozen on frame balance, with the weak anomaly prior (worst-KS ep 3, offset-mean ep 4) used only to pick which half trains first.

Dataset names pinned by holdout-draw search (the gripfix Amendment-1 class, pre-applied; holdout_episodes is a pure function of the repo name):

• Branch A: mcobzarenco/so101_pick_place_clean_ch0fix_act_j — 7 episodes, draw (2,): the clean-side train split stays episode-identical to democlean’s ({0,1,3,4,5,6}, 3026 kept frames).
• Branch B cell: mcobzarenco/so101_pick_place_clean_ep015_c — 4 episodes (originals [0,1,2,5] in ascending order, re-indexed 0–3), draw (2,) = original episode 2, the decoy: episode 2 is democlean’s own never-trained holdout, carried in each subset precisely so the mandatory ≥1-episode holdout lands on it and the cell trains its full intended subset {0,1,5} while preserving “episode 2 never trains” semantics.
• Branch B registered follow-up: mcobzarenco/so101_pick_place_clean_ep346_a — originals [2,3,4,6], draw (0,) = original episode 2, same decoy design, trains exactly {3,4,6}.

Branch selection (frozen; no judgment at the verdict)

The rung-2 grid verdict on pdnorm_ch0fix’s sim100 battery is the selector: ≥20/100 → branch A. ≤10/100 → branch B. 11–19 → neither (rung 2’s ambiguous-band protocol runs; any rung-3 launch would then need a fresh registered amendment). The verdict session’s only jobs are: bank the rung-2 verdict per its own pre-reg, build the selected branch’s materializer, run its oracles, fit-smoke, launch, announce.

Branch A — the action-only cell (fires iff ch0fix ≥ 20/100)

Question: rung 2’s affine touched ch0 in BOTH the action and state columns. Which half carries the recovery? The queue item named two decomposition axes; one is already settled at zero cost — the banked constant-freeze read pre-refuted the shift form (KS unmoved 0.295→0.286/0.308) and the shift component of the winning affine is 0.05 of demos’ std, so scale-vs-shift is closed: the scale is the affine. The open axis is action vs state, and it is the axis with production value: an action-only fix is output-side data hygiene any pipeline could apply mechanically; a state-side requirement means the fix must stay consistent with what the policy is told about the world — a stronger, more invasive rule.

The cell. Mix = grasp_demos_v2/merged + so101_pick_place_clean_ch0fix_act_j (×4), and nothing else — the 7 clean episodes with the ch0 action column only transformed by the rung-2 frozen affine, constants verbatim:

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

The ch0 state column stays byte-identical to source (that is the treatment), as does every other channel, count, camera stream, and annotation. --recompute-stats gives the set its own pdnorm rows — the action row’s ch0 scale must move ×2.7552 while the state row’s stays clean-like, which doubles as the live oracle that exactly one column was touched.

Frozen transform decisions: the rung-2 affine exactly, not re-estimated; ch0 only; action column only; float64 transform cast back to source dtype.

Materializer: fontaine/scripts/make_clean_ch0fix_act_dataset.py (exec session; sharing code with make_clean_ch0fix_dataset.py is a file-format freedom). Oracles: action ch0 equals the affine of source exactly; state ch0 byte-identical to source; every non-ch0 column byte-identical; frame/episode counts identical; transformed action range inside demos’ observed support [−110.0, +79.6]; holdout draw is (2,); the rung-2 no-op guard (refuse if source ch0 std exceeds 20).

Reads and grid (primary sim100 flow leg at step 3000, protocol byte-identical to the five banked batteries):

• ≥ 20/100 → the action column alone carries: the mix-hygiene rule is output-side and mechanically deployable; state-side consistency is not required. Mechanism cell closes the ch0 thread; no further rung named.
• ≤ 10/100 → the joint edit was necessary — the recovery needs the state column too. Named follow-up (delegation standing): the state-only complement cell, own draft, same discipline; the mechanism claim is capped per the honesty clause below.
• 11–19 → ambiguous (the control’s band): no claim.

Paired reads recorded alongside: vs ch0fix’s own banked number (THE read — same 7 episodes, the both-columns twin), vs democlean 8/100, vs onerig 28/100, vs control 11/100.

Registered asymmetry note (honesty clause). The action-only edit deliberately manufactures a within-dataset action/state inconsistency on ch0: the model trains on frames whose state says “compressed sweep” while the action commands the stretched one. A ≤10 outcome therefore cannot fully distinguish “the state half carries” from “the inconsistency itself is an artifact that re-poisons” — the claim caps at “action-only is not a sufficient fix; the joint edit stands as the deployable form”. The paired-progress read (the gripfix −2.07 cm precedent) carries artifact detection.

Branch B — the content-bisection cell (fires iff ch0fix ≤ 10/100)

Question: with the gripper amplitude struck at rung 1 and the ch0 marginal struck at rung 2, the manifold probe’s named-suspect list is exhausted — the carrier is content-level. Rung 2’s ≤10 branch mandated a design pass before committing to per-episode leave-one-out (~17 GPU-h × 7 ≈ 119 GPU-h). This is that design pass, and the verdict is bisection: one cell per rung halves the candidate set at the standard cell price; a localized carrier is convicted to a single episode in ≤3 rungs (~51 GPU-h worst case), and a distributed carrier — which per-episode LOO cannot even represent — is detected in two.

The cell. Mix = grasp_demos_v2/merged + so101_pick_place_clean_ep015_c (×4), and nothing else — episodes {0,1,5} of clean plus decoy episode 2, every kept frame byte-identical to source (no value edits of any kind; the treatment is subset membership). The trained clean content is exactly episodes {0,1,5} (1,504 frames; democlean trains these plus {3,4,6}’s 1,522).

Frozen split decisions: the six trained episodes split frame-balanced — {0,1,5} (1,504) vs {3,4,6} (1,522); no measured channel ranks the episodes (basis table above), so balance is the criterion. The first cell trains {0,1,5} and drops {3,4,6}, the half holding the two most marginal-anomalous episodes (ep 3: worst per-episode KS 0.469, +12.3 mean offset; ep 4: −7.1 mean offset) — under the weak suspects prior this maximizes the chance of the decisive outcome (recovery ⇒ carrier localized in the dropped half). Decoy-holdout design as pinned above; episodes re-indexed 0..3 in ascending original order.

Materializer: fontaine/scripts/make_clean_subset_dataset.py (exec session; takes the episode list + decoy). Oracles: episode set exactly the pinned originals in ascending order; every kept frame’s every column byte-identical to source; per-episode frame counts match the basis table; holdout draw (2,) for ep015_c / (0,) for ep346_a; the --dataset-repeat 'mcobzarenco/so101_pick_place*=4' glob matches the new name.

Registered dose confound (honesty clause, priced in before launch). The cell trains 1,504 clean frames where democlean trains 3,026 — effective share ~0.34% vs 0.69% at the verbatim ×4 repeat (one-dataset-delta discipline; no repeat-flag compensation, that would be a second treatment variable). A recovery is therefore ambiguous between “carrier in the dropped half” and “half dose is below the poisoning threshold” — rung 2 of this sub-ladder, the complement cell ep346_a, is the registered follow-up on EITHER verdict and disambiguates:

• cell recovers, complement collapses → carrier localized in {3,4,6}; recurse there.
• cell collapses (poison survives at half dose in {0,1,5}) → complement next: if it ALSO collapses, the carrier is distributed — bisection and LOO are both moot at this grain; escalate to the owner with content forensics, no auto-launch beyond the complement.
• both recover → dose/dilution, not localized content — a different, cheaper mechanism class (share-threshold sweep), own pre-reg.

Reads and grid (primary sim100 at 3000, protocol byte-identical):

• ≥ 20/100 → grasping recovers without {3,4,6} at half dose; complement cell next (localization vs dose per the table above).
• ≤ 10/100 → the poison rides {0,1,5} at half dose; complement cell next (distributedness test before any recursion).
• 11–19 → ambiguous: no claim, complement decision escalates to the owner.

Paired reads recorded alongside: vs democlean 8/100 (THE read — same recipe, content subset), vs onerig 28/100, vs control 11/100.

Shared spec (both branches)

Command: the democlean launcher verbatim (launch_local_grasp_sft_v2_joint_1gpu_pdnorm_democlean_h100.sh) with exactly one delta: the dataset name in --train-data (the repeat 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 six anchors). Fit smoke first; compute-app abort guard + policy-server check carried.

Anchors (all banked): ch0fix (the rung-2 number, whichever band it lands in), democlean 8/100, gripfix 5/100, onerig 28/100, control 11/100, convicted 1/100.

Secondary — drift guard: Δeval(1000−500) ≤ +0.30, record-only (the probe cannot clear these cells — decoupling banked four 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 five cells.

Record-only: the recomputed pdnorm rows (branch A: action-row ch0 scale ×2.7552 with the state row pinned ≈ clean’s — the one-column-touched oracle; branch B: subset rows vs clean’s banked rows, no edit expected); eval-250 curve vs the democlean curve.

Gates and boundaries

• GPU-hours gate: 17 per fired cell (train ~13.7 measured on the identical recipe + battery ~3); exactly one cell fires from this draft.
• Launch window: the ch0fix verdict session, immediately after the rung-2 verdict banks — delegated, announced in-channel, never gated on an owner GO. The H100 is free at that boundary (ch0fix train + battery complete).
• 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 branch grid.
• Checkpoint policy: saves under ~/checkpoints/finetune/grasp_sft_v2_joint_1gpu_pdnorm_ch0fix_act (A) / ..._pdnorm_ep015 (B); endpoint banks weights-only on any decisive read.
• Seed policy: seed 0 (comparability); ×4 repeat carried unchanged; branch B’s share change is the registered confound above, not a knob.

Objection/decision path: branch selection is mechanical from the rung-2 grid; reads, grids, split, affine constants, and dataset names frozen as drafted. The exec session’s only freedoms are materializer file-format details and the in-window launch timing. Any spec change lands as a registered amendment first.