Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Pre-registration: state-dropout 0.8 — the anti-shortcut arm (ideas #9)

2026-08-06 ~08:1xZ. Posted before launch (charter §4). The state-reliance probe’s branch rule fired (results: D = +0.702 [0.498, 0.916] — aux-off leans harder on the state shortcut, and all four checkpoints degrade +15–16 chunk MAE under state masking), so the literature’s counter-lever gets its paired arm: Adapt Your Body-style state dropout at p = 0.8, the causal-confusion line’s standard mitigation (2509.18644 goes to state-free entirely and reports better spatial generalization). Explore-class arm (charter §3 budget): the modal outcome is “within band”; the tail is a vision-reliant policy, which is what the §0 north star (few-shot rig transfer) actually needs — a policy that extrapolates proprioception transfers worse than one that looks.

Instrument (landed with this post, before launch)

--state-dropout p in bijou.train: with probability p per sample, the collator rewrites the item through mask_state_item — the SAME primitive as the eval probe’s --mask-state (state → dataset mean ⇒ normalized soft state token EXACTLY zero, raw decoder-side state at the mean, actions/targets untouched), so the regularizer trains exactly the condition the probe measures and the two can never drift. p=0 is bitwise-inert and consumes no RNG (all three CPU loss oracles reproduce exactly: flow 2.7903/1.9152, ar_fast 4.9232/4.8631, ar_backbone 27.8262/27.7701; check.py 240 green; SnapFlow stage-0 recipe verify re-run green with the new field at its inert default). In-run probes and panel evals score INTACT state (dropout-0 clones — the deployment condition); the masked readout is the offline probe.

Design — one arm, one variable

Arm C: fontaine_arb_rcond_statedrop80_40k_1xh100 — the A-s0 recipe verbatim (seed 0, 40k steps, B10, ar_backbone rcond, box GPU 0, launcher fontaine/scripts/box/launch_box_gpu0_fontaine_arb_rcond_statedrop80_40k_1xh100.sh) plus exactly --state-dropout 0.8 (the literature’s value; p is the one free parameter and 0.8 is the aggressive end — if it fails while the mechanism reads clean, a lower-p screen is the pre-declared branch, a NEW pre-reg). Paired against A-s0’s banked panel npz (chunk 7.7966 / first 3.9422 @40k) — zero re-evals on the intact side. Chained in-launcher after training: full panel eval, then the masked-subset reliance eval (frozen q4 plan, sha256 asserted at launch before any GPU work).

Known seam, stated now: the per-sample dropout draw comes from the same per-worker generator as the other train-time regularizers, so arm C’s camera-kind/instruction/condition dropout decisions shift relative to A-s0’s (equivalent to re-seeding the regularizer streams only; data order, model init, and τ/ε draws are untouched — separate generators). This sub-seed effect is bounded above by the full seed-noise measurement: max pairwise replicate delta 0.0697, σ_seed 0.038. The 0.15 band floor already absorbs it.

Frozen reads

  1. Primary — paired per-frame panel chunk_mae, C − A-s0 (17,204 core frames, seeded bootstrap CI, the box-batch instrument’s paired-read path re-oracled on the new pair before the real read). Band = max(3σ_seed, 0.15) = 0.15:
    • C − A-s0 < −0.15 ⇒ state dropout HELPS actions in-distribution → adopt as a recipe default in future arms (its own follow-up pre-reg for the next lineage run).
    • |C − A-s0| ≤ 0.15 ⇒ headline-neutral → decision moves to reads 2–3: a free hardening lever is still adopted if the mechanism reads clean (vision reliance is north-star-relevant even when the in-distribution panel can’t see it).
    • C − A-s0 > +0.15 ⇒ p=0.8 costs actions → adopt NOTHING; branch: if read 2 shows the mechanism worked (reliance collapsed), a p=0.3 screen is the one sanctioned follow-up; if read 2 also fails, #9’s dropout leg is falsified at this scale.
  2. Reliance readout — masked-subset eval of C @40k (frozen q4 plan, 4,301 rows; intact side pooled from C’s own panel npz — strict row-subset, the probe instrument’s established pattern). Execution oracles inherited from the probe pre-reg: state-copy/ -norm rows byte-match C’s own full-panel npz on the subset rows; report JSON records mask_state: true; policy name carries _state-masked. Reads, with banked comparators:
    • Sanity gate: Δ_first(C) = masked − intact first_mae < 5.0 (vs A-s0’s +19.950): masking is in-distribution for C by construction — this gate only proves the regularizer trained the intended condition; it is NOT evidence for the vision-shift hypothesis.
    • The capability number: C’s masked first_mae — vision-only first-frame prediction. Banked comparators on the identical rows: intact state-copy 2.4316 (the proprioceptive-extrapolation floor), A-s0 masked 23.8154, B masked 24.0783. Masked-C beating state-copy (< 2.43) would be a qualitative first (no existing arm is within 20 of it); pre-declared reporting thresholds: < 6.0 = strong vision capability signal, ≥ 15 = dropout failed to build one.
  3. Grounding read — C’s intact panel first_mae vs A-s0’s 3.9422 (the grounding-sensitive column; state-copy 2.6202 floor). Interpretive trap pre-declared: B (aux-off) IMPROVED first_mae (3.5009) by leaning on state — read 3 is only meaningful jointly with read 2 (better first_mae + collapsed reliance = vision did it; better first_mae + intact reliance = shortcut did it).

Expectations (banked before launch)

  • E1 startup (hard gate): 878 datasets / 42,872 episodes / dims 6/6, identical to the batch; banner prints state dropout: p=0.8; bijou_config.json records state_dropout: 0.8. Any deviation ⇒ abort before step 1.
  • E2 throughput: 0.4–0.6 s/step at B10 (idle box, no contention); sustained > 0.8 ⇒ input fix at a save boundary. VRAM < 76 GiB (state masking changes no tensor shapes; memory must match A-s0’s profile).
  • E3 in-run probe (intact state, ±0.3 floor): expect a trajectory ≤ ~1 above A-s0’s (6.955@39k) — 80% of samples lose an informative input, some convergence cost is honest to expect. Soft: < 13 @10k, < 10.5 @30k. Kill: > 13 @10k after falling-then-rising; NaN; second OOM after the standing B−1 resume. Formal final gate: probe < 10 @40k — above that, p=0.8 is too aggressive at this scale and reads 1–3 still run (a negative is a deliverable), but no adoption path opens from this arm.
  • E4 panel (the honest prior): chunk_mae C in 7.65–8.30 — centered slightly worse than A-s0 (information removed at train time; the literature’s wins are OOD/spatial-generalization wins, which this in-distribution panel largely cannot see). first_mae anywhere in 3.4–4.4 given the read-3 trap. The pre-registered question is WHERE in the band it lands and what read 2 says, not whether C beats the baseline.
  • E5 reliance: Δ_first(C) < 5.0 (sanity); masked-C first_mae is reported against the three banked comparators above, whatever it is.

Cost & environment

~5.0 h train (40k × ~0.45 s/step, 1×H100, idle box GPU 0) + ~2.5–3.3 h panel eval + ~25 min masked eval ≈ 8.5 GPU-h, on a box otherwise idle awaiting owner steer (the posted E4B follow-on recommendation was “box to grounding arms” — this IS a grounding arm, from the same mechanism chain). Disk: one run × 8 saves × 24G = 192G into 6.8T free. Same corpus, same selection, no derived data ⇒ no new leakage surface (E1 asserts selection identity with the batch). Box code fetched to the commit carrying this post before launch (box idle — no live-run sync); wandb fontaine, run name = save dir name, one lineage.

Falsification honesty: this arm cannot distinguish “state dropout builds vision grounding” from “state dropout merely tolerates masking” if read 1 is neutral AND read 2’s capability number stays near the 20s. That combination — trained-in mask tolerance without transfer to intact-state grounding — is the pre-declared “mechanism inert” outcome, and it kills the dropout leg as cleanly as a band miss.