Pre-registration DRAFT: the Squint-twin qualification screen — first use of the SO-101 digital twin, relative reads only

Draft cut 2026-08-22 00:2xZ (work session riding the gripfix endpoint battery), queue item squint-twin-screen-prereg — the eval-design v0’s named slot 2. NO launch from this item. DRAFT status: the reads, gates, and expectation grid below are frozen now and do not move; the finalization amendment (before any launch, in-channel per the standing no-GO-ask rule) fills exactly the named slots — the preflight-2 receipts (§7), the adaptation-recipe command block, the frozen instruction strings, and the pair-2 slot that waits on tonight’s gripfix frozen-grid verdict. Substrate: Squint (MIT, local checkout ~/squint), preflighted GO 2026-08-14 (note).

Plain words. We have a second simulator available: a digital twin of exactly our robot arm, published by another lab, with automatic success grading. Before ever letting it influence a decision, we have to qualify it — the way you’d qualify a new measuring instrument against one you already trust. The plan has three gates. First, mechanical: prove our policies’ outputs actually drive the twin correctly (units, joint order, camera plumbing), by replaying a real robot episode inside it. Second, a positive control: teach one of our models the twin’s alien-looking world with a small dose of in-twin demonstrations — if even an adapted model can’t score, the twin can’t referee anything and we close the screen as a broken instrument. Third, the actual measurement: take two of our checkpoints whose quality difference we know precisely from our own simulator (28/100 vs 8/100 at grasping), adapt both identically, and ask whether the twin sees the same ordering. If it does, we’ve bought an independent second opinion for future model comparisons — plus an unlimited supply of ground-truth-labeled robot videos for calibrating failure detectors. If it doesn’t, we record that and the twin stays a toy. Absolute scores in the twin are never claims — only differences between our own models, measured under identical conditions.

The tier decision, resolved (the item’s first deliverable)

The preflight note’s rule (2026-08-14): “Squint is the successor tier if the wrist-transfer screen hits F-instrument or the success floor holds.” Both branches have since fired:

• The wrist-transfer screen closed F-instrument (08-15): its control couldn’t certify its own instrument at n=25, and the success-rate form of the transfer question was left standing.
• Its successor-requirement #1 — “a competence floor first: re-screen on a policy whose success rate can actually move” — is now met: the grasp-SFT lineage has a 28/100 checkpoint (onerig) and a certified 20-point paired spread below it.

Tier decision: GO for qualification. Not GO for judging: per eval-design v0, the twin’s registered role until this screen passes is none — it screens nothing, gates nothing. This pre-registration IS the gate before first verdict-adjacent use.

The question and the registered role

One question: can the Squint twin, with the domain gap held constant across arms, reproduce a capability ordering that sim100 certified? Pass → the twin registers in eval-design v1 as a relative screening instrument (A/B deltas between our checkpoints, never absolutes, never a verdict-gater). Fail → it registers as a labeled-rollout generator only (idea #6), and the menu’s next tier stays sim100-only.

Gate 0 — mechanical adapter (CPU, no policy)

The adapter (a ~100-line gym loop per the preflight) carries four frozen conversions:

• Units + joint order: policy outputs LeRobot servo degrees; the twin’s pd_joint_pos consumes absolute radians (normalize_action=False, bounds = joint limits). Arm joints: deg2rad. Gripper: the twin’s own deploy path encodes a gripper-specific affine (sim convention −10°..120° ↔ servo open constant; deploy_utils/manipulator.py:49-52,135) — we invert it with OUR calibration’s open value (demos command exactly {0, 41.69}); the two affine constants are a preflight-2 receipt.
• Rate: our chunks are 30 Hz (sim100 CONTROL_HZ = 30); the twin is hard-pinned at control_freq=10 (envs/base_random_env.py:148). Frozen: subsample every 3rd action, execute 10 env-steps per replan — 1 s wall-equivalent per replan, the same cadence as sim100’s execute-horizon 30. The published twin dynamics are not touched (no SimConfig overrides).
• Cameras: the dual-camera subclass the preflight priced (wrist mount + third mount both exist in the base classes), 224×224 via sensor_configs, apply_overlay=False (raw renders — the rgb+segmentation trap is moot), domain_randomization=False (deterministic paired seeds, verified in preflight). Wrist frame → [wrist camera|…]; the third-person frame’s kind tag is judged by resolve_camera_kinds on the saved preflight frame — a preflight-2 receipt, frozen at finalization, identical across arms.
• Decode + stats: sim100 protocol constants verbatim — euler-10, bf16 decoder, each checkpoint’s own training stats (grasp_demos_v2/merged worn-row rule carried).

Gate 0 read (pass/fail): (a) determinism entry gate — one seed rolled twice, per-step obs bit-equal (wrist-screen precedent); (b) rig-episode replay — a banked grasp_demos_v2 episode’s action trace driven through the adapter into the twin: tracking p50 error < 0.05 rad across arm joints, no joint-limit violations, gripper open/close events reproduced. Fail → fix or close; nothing downstream launches on a failed adapter.

Gate 1 — the sim-adaptation sanity arm (the positive control)

Separates “renderer is alien” from “policy is bad”. The wrist screen’s successor rule applies: the control runs at treatment n (n=100), not a cut-rate pilot — a control that can’t fire is decoration.

• Demo source: the repo ships no policy checkpoints — we train Squint’s own state-based SAC expert per task (train_squint.py, 2–9 min/task on a 3090-class GPU; priced ≤0.5 GPU-h for the 2-task set), then roll it out and keep ~100 success episodes/task, re-rendered through OUR adapter config (dual-camera 224, raw, DR off) — the twin gives ground-truth success labels for free.
• Conversion: twin episodes → LeRobot format (actions/state back in servo-degree convention through the inverse adapter — the same constants Gate 0 verified). Conversion oracle: one episode round-tripped frame- and action-bit-exact.
• Adaptation: short SFT via bijou.train (the only training path), starting from each arm’s checkpoint, one frozen recipe, identical across arms — same steps, LR, seed, data; no tuning, no second attempt. Budget hard cap ≤2.5 GPU-h/arm; exact step count in the finalization command block.
• Gate 1 read (pass/fail): the adapted stronger arm scores ≥20/100 on its best task after the band pilot (below). Fail → F-instrument: the twin cannot be made legible to our stack at this budget; screen closes, no relative read is attempted, the labeled-rollout deliverable (#6) survives.

Gate 2 — the qualification read (the screen proper)

• Pair 1 (frozen now, verdict-independent): grasp_sft_v2_joint_1gpu_pdnorm_onerig @3000 (28/100) vs …_democlean @3000 (8/100) — the maximum-contrast banked pair, 20 points on paired seeds, and (pending tonight’s gripfix verdict) a gripper-amplitude mechanism that Lift/Place tasks exercise directly. Pair-2 slot: filled at finalization from the gripfix frozen-grid verdict (≥20 → gripfix joins as the hygiene-rule arm); pair 2 is optional and never blocks pair 1.
• Tasks: SO101LiftCube-v1 primary, SO101PlaceCube-v1 secondary — piloted by the Gate-1 adapted arm into the 20–80% band (the banked bench rule); out-of-band → substitution ladder Reach (easier) / Stack (harder), frozen order, logged.
• Cells: n=100 paired seeds (seeds 0–99), DR off, both arms adapted by the identical Gate-1 recipe. The unadapted pair runs the same cells as a free record-only rider (expected at/near floor; if it isn’t, that’s worth knowing and costs nothing extra).
• Primary read: paired per-seed McNemar on success, adapted onerig vs adapted democlean, primary task — the standing sim100 machinery (sim100_paired_read.py class).
• Co-primary (rig-grammar rehearsal): per-episode time-to-predicate curves from the twin’s honest info predicates (reached_object → is_item_grasped → item_lifted), Kaplan–Meier with timeouts right-censored + KS with episode-clustered bootstrap — the eval-design tier-3 analysis run for the first time on cheap rollouts. Full CDF panel published; no scalar summaries without it.

Frozen expectation grid:

Budget and gates

SAC experts ~0.5 + adaptation ≤2.5×2 + rollout inference (~800 episodes × ~5 replans × ~0.6 s + pilots) ~1 + margin ≈ cell gate ≤7 GPU-h, abort-and-close at the gate. Sim overhead itself is noise (1.35 s/ep at the CPU floor). Every GPU leg launches systemd-run detached at a free window, policy-server checked, per the standing rules; the twin never preempts a sim100 verdict leg.

Preflight-2 receipts (CPU, before finalization)

1. Dual-camera subclass renders both views in one env (saved frames).
2. resolve_camera_kinds judgment on the third-person frame → frozen kind tag.
3. Gripper affine constants extracted (ours: open 41.69; twin sim convention −10..120) + Gate-0 replay green.
4. LeRobot conversion oracle green on one round-tripped episode.
5. train_squint.py smoke (imports + a few CPU env steps; no GPU).

Claims contract (frozen)

1. Relative deltas only, domain gap held constant across arms.
2. Absolute twin numbers are never claims — the paper’s own in-domain BC baseline at 41.9% sim success is the permanent warning label; ours sit under a far larger gap.
3. No sim-to-real claims — the twin is visually far-OOD from both the rig and sim100; nothing here measures transfer.
4. One role per instrument: this screen can promote the twin to relative screening instrument, nothing higher; twin reads never gate a sim100 verdict, and a passed screen is not license to skip sim100 — it adds a second opinion, it replaces nothing.
5. Silence proves nothing: a null Gate-2 read demotes the twin, it does not exonerate either checkpoint.

What this feeds

Eval-design v1 slot 2 (this pre-reg), idea #16 (the substrate menu: sim100 / twin / rollout-free / rig), idea #6 (the labeled-rollout corpus: every twin episode banks frames + per-step predicates + ground-truth success — detector calibration data at unlimited volume).

Appendix (2026-08-22 00:4xZ, same session): preflight-2 receipts — EXECUTED

The §“Preflight-2 receipts” list, run CPU-only in the twin’s venv while the gripfix battery rode gpu0 (CUDA masked + lavapipe, GPU untouched). Probe: fontaine/scripts/squint_preflight2.py; facts: outputs/squint_preflight2/facts.json (mirrored to fontaine-reports). These are receipts, not results — the finalization amendment freezes the two re-priced lines below before any launch.

• R1 dual-camera subclass: GREEN. One env serves both views at (1, 224, 224, 3) — the wrist camera inherited (mount + per-step pose update), a static third_camera added on the base env’s existing camera_mount with ThirdCameraEnv’s published pose/FOV constants. ~25 lines, as priced. Wrist · third. Kind-tag observation for the frozen slot: the third view is an elevated three-quarter view (eye 0.25 m up, ~40° down), not an overhead — the honest vocabulary choice is front (or side), not top; our checkpoints trained on top+wrist only, so the tag is OOD either way. Frozen at finalization.
• R2 determinism entry gate: GREEN. Same seed + same 10-step action sequence twice → sensor bytes and qpos bit-equal (DR off).
• R3 Gate-0 rig-episode replay: tracking GREEN, limit line needs the re-price. grasp_demos_v2/merged ep 0 (449 frames @30 Hz → 150 twin steps @10 Hz) through the frozen mapping: tracking p50 0.0025 rad (gate <0.05 — 20× under), p95 0.148 (fast-transient lag, consistent with 10 Hz steps over a 30 Hz path), gripper transitions 3/3 commanded→achieved. But the drafted “zero limit violations” line is unmeetable as written: 41 clips, all ≤0.047 rad — forensics: 36× wrist_flex grazing its limit by ≤0.0032 rad, 3× shoulder_lift at ≤0.0471 rad (the twin’s −1.7453 rad limit sits ~2.7° inside our demo’s deepest reach −1.792), 2× elbow_flex ≤0.003. A banked substrate fact, not an adapter bug. Finalization re-price (named slot): Gate-0 limit line becomes “no arm clip > 0.05 rad” — preserving the intent (catch unit/order/sign errors, which produce radian-scale clips) without failing on the twin’s genuinely tighter joint limits.
• R4 train_squint.py smoke: GREEN after dep install (wandb, tensordict, torchrl, tqdm, torchvision added to the isolated venv — the 08-14 preflight installed eval deps only). Recorded venv drift: the torchrl install upgraded the venv’s torch 2.6.0-cpu → 2.13.0+cu130. All receipts re-taken under the final stack; the replay statistics are bit-identical across the torch change (CPU PhysX physics unmoved). The exec session inherits the venv as-is; CUDA_VISIBLE_DEVICES masking is what keeps CPU legs off the GPU now that the build can see it.

Open before FINAL: the LeRobot conversion oracle (needs the conversion tooling the exec item builds), the adaptation-recipe command block, instruction strings, the third-camera kind tag, the Gate-0 limit-line re-price above, and the pair-2 slot (gripfix verdict, tonight).

FINALIZATION AMENDMENT (2026-08-22 02:5xZ) — all named slots frozen; launch follows this post

Exec session squint-twin-screen-exec part (b), announced in-channel per the standing no-GO-ask rule. Every open slot above is resolved here; nothing else in the pre-registration moves. The pipeline below was smoke-tested end-to-end (junk 30k-step expert → rollout → re-render → LeRobot → oracle) before this amendment; smoke artifacts deleted, receipts quoted where they price a frozen line.

Slot 1 — Gate-0 limit line (re-priced as named): “no arm clip

0.05 rad“. R3 stands green under this line (41 clips, all ≤0.047 rad, twin-limit geometry not adapter error).

Slot 2 — third-camera kind tag: front. Dataset key observation.images.front, mirroring grasp_demos_v2’s own key pair — the twin dataset wears exactly {wrist, front} like the rig demos (twin base_camera → wrist, third_camera → front).

Slot 3 — instruction strings (frozen):

• LiftCube: Pick up the red cube.
• PlaceCube: Pick up the red cube and place it in the bin.

Slot 4 — conversion spec + oracle re-price. Twin episodes (10 Hz) → LeRobot v3 at fps 30 by repeat-3 upsampling, so the deploy adapter’s subsample-every-3rd exactly inverts the conversion. Actions = the rollout controller’s absolute radian targets (target-delta control is PDJointPosController with use_target/use_delta — _target_qpos IS the pd_joint_pos-equivalent command), inverse-adapted to servo degrees; state likewise from qpos. Oracle re-price: the drafted “action-bit-exact” becomes round-trip max |Δ| < 1e-5 rad (deg↔rad float paths aren’t bit-stable; smoke measured 6.6e-8), frames bit-exact at the pre-encode boundary, and decoded-video PSNR recorded as a fact, not a gate — our own demos ride the same mp4 path (smoke: ~36–40 dB). Dataset: ~/datasets/fontaine/squint_twin_demos_v1, ≤100 episodes/task.

Slot 5 — expert + collection config (frozen). Experts: train_squint.py paper defaults with --no-env-domain-randomization (DR off everywhere in this screen), seed 1, no wandb, ckpt = last 100k-step eval save, hard cap 45 min/task. Collection: rollout in the expert’s native env config; re-render through the dual-camera 224 raw DR-off pd_joint_pos env, same seed + set_state_dict; the final success label is the re-render’s own end-state success (episodes losing success in replay are dropped; keep-rate + replay-divergence receipts banked — smoke divergence p50 0.0038 rad). Per-step honest predicates banked per episode (the idea-#6 corpus starts here).

Slot 6 — adaptation recipe command block (frozen, identical both arms). Priced from the pdnorm lineage’s measured 16.2 s/step at eff-96: 500 steps ≈ 2.25 GPU-h/arm ≤ the 2.5 cap.

uv run python -m bijou.train \
  --train-data ~/datasets/fontaine/squint_twin_demos_v1 \
  --init-from ~/checkpoints/finetune/grasp_sft_v2_joint_1gpu_pdnorm_{onerig|democlean}/step_003000 \
  --objective joint --joint-ce-weight 1.0 --insulate-flow \
  --recompute-stats --per-dataset-flow-norm \
  --flow-decoder-init inherit --image-augment 0.8 \
  --decoder-lr 5e-5 --backbone-text-lr 1e-5 \
  --steps 500 --batch-size 96 --backward-chunks 8 \
  --activation-checkpointing --offload-optim --prune-superseded-optim \
  --holdout-episodes 0.1 --eval-every 100 --eval-samples 256 \
  --save-every 250

Adapted arms wear their own recomputed stats (the twin rows); the unadapted record-only rider wears the sim100 worn-row rule verbatim.

Slot 7 — pair-2: EMPTY. The gripfix frozen-grid verdict landed 5/100 (≤10 band); the frozen rule (“≥20 → gripfix joins”) does not fire. The screen runs pair 1 only.

Launch plan: leg A now (experts → collection → conversion → oracle, one detached unit, ~0.8 GPU-h); leg B (adaptation, both arms chained, ~4.5 GPU-h) after the oracle receipt is verified; Gate-2 rollout harness at the next window. Cell gate ≤7 GPU-h unchanged.

RESULTS (2026-08-22 13:57Z; appended 2026-08-22 23:0xZ work session): Gate 1 FAILED — screen closed F_INSTRUMENT

Full record: results page. Verdict artifact outputs/squint_screen/eval/gate1.log.

• Gate 0: PASS (preflight-2 receipts + leg A oracle — replay tracking p50 0.0025 rad, no arm clip > 0.05 rad, round-trip 6.5e-8 rad, frames bit-exact pre-encode).
• Leg A COMPLETE 05:43Z (~1.9 GPU-h): experts success 1.00 both tasks; squint_twin_demos_v1 banked, 100+100 episodes with per-step ground-truth predicates.
• Leg B COMPLETE 13:33Z (r4 full retrain ~4.4 GPU-h after the ENOSPC + resume-bug incident, ~2.7 GPU-h re-spend, crossing recorded in-channel): frozen Slot-6 recipe both arms, probe twins onerig 2.47@500 / democlean 2.5187@500.
• Gate 1: FAIL. Adapted onerig @500: band pilots 0/20 both tasks (BELOW_BAND), cells completed to n=100 anyway — 0/100 lift, 0/100 place vs the ≥20/100 best-task bar → FAIL_F_INSTRUMENT, the pre-registered valid end. Milestone forensics: reached_object 20/100, grasps 3–4/100, item_lifted 7–10/100, success 0 — partial competence, not a transport flatline. No relative read attempted; Gate-2 spend skipped; substitution ladder (Reach) logged, never auto-run — a future pre-registered session’s call.
• Registered role per the frozen fail branch: the twin is a labeled-rollout generator only (idea #6 — the corpus survives and banked its first 400 predicate-labeled episodes). It screens nothing, gates nothing; the next eval tier stays sim100-only. onerig 28/100 vs democlean 8/100 stands exactly where sim100 certified it.