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-reg (DRAFT) — GRPO R2 on the grasp-SFT policy: RL pressure on a competent base

2026-08-15, drafted ~07:5xZ while stage C of the grasp-SFT bootstrap trains (313-demo set, AR primary, launched 07:29:55Z). Status: DRAFT, conditional — this registration activates ONLY if stage D’s frozen primary reads GRPO GO (≥ 20/100); a 5–19 read routes to the bootstrap’s own iterate-once arm first, and < 5 parks this page (F-transfer makes visual work the binding lane, not RL). Finalization (frozen params + objection window + HEAD re-pin) happens AFTER the stage-D verdict banks, per Decision 11: any new GRPO run is a fresh pre-reg on the first-class stack. Nothing here launches before that.

Plain words. We stopped our last reinforcement-learning run with a verdict worth restating: the robot was too incompetent for incentives to matter. It succeeded so rarely (roughly 1 try in 20 at best) that the training signal was mostly noise about shoving, and even after we fixed the scoring rule so shoving never pays, the run went nowhere — the conclusion was “teach the robot to grasp first, then come back.” That teaching is happening right now: a scripted expert generated 313 successful demonstrations, and a policy is being fine-tuned on them as this draft is written. If that policy passes its exam (at least 20 successes on the 100 held-out test scenarios), reinforcement learning finally has something to work with: when the robot succeeds 1 time in 5 rather than 1 in 100, every batch of attempts contains both successes and failures to compare, which is exactly the contrast this kind of training amplifies. This document pre-commits how that run would work — same scoring rule, same safety wires, measured budget — so the go/no-go decision is mechanical when the exam result lands.

§1 Hypothesis — why R2 stops being moot

The R1-B boundary read closed phase 2 on surface A with “the R2 pricing discussion is moot” — at that policy’s competence (successes 4/3/3 of 64 per wave; held-out 2/20 unchanged). The registered next shape was grasp-rich SFT before RL pressure, and the bootstrap pre-reg §4 F-live froze the return path: ≥ 20/100 → GRPO registers fresh with success-variance groups.

H: on a base policy at success rate p ≥ 0.20, GRPO’s group advantages are carried by the success term (+10, the largest reward magnitude) rather than by centimeter-scale progress noise, and gentle re-pricing measurably accumulates held-out success where R1-A/B measured flat.

The arithmetic of the claim: a group of 8 draws on one seed is informative when it mixes successes and failures — probability 1 − (1−p)⁸ − p⁸. At R1-B’s measured competence (p ≈ 0.05) that is ~34%, and the mixed groups’ contrast was a ±10 spike on top of a reward otherwise dominated by shove/progress terms the policy could not control. At the stage-D gate minimum (p = 0.20) it is ~83% of groups, each carrying a ±10 contrast between trajectories the base policy itself produced. Same loop, same reward — the signal changes class because the base changed class.

§2 Run design (frozen at draft; finalization fills §6)

  • Stack (Decision 11): the first-class new stack — bijou/grpo_replay.py / MolmoAct2DiscreteStack over bijou checkpoints. FRESH run; the banked R1-A/B .pt endpoints stay salvage-only, nothing resumes across the retirement re-point. Full-width Gumbel noted: sampled streams are not comparable draw-for-draw with the banked port waves.
  • Base policy: the stage-C AR endpoint through the same two-hop conversion stage D evaluates (bijou.convert_molmoact2, demo-set recomputed norm_stats riding the HF dir, no shim per bootstrap §6) — GRPO starts from byte-identically the policy whose 20+/100 was measured. Checkpoint id + conversion receipt pinned at finalization.
  • Reward: composite_reward_v2 trained-on (grasped-progress only, 0.5 ungrasped charge direction-blind, +10 success −2 tipped −5 strike), held-out metric stays v1 — both unchanged from R1-B, keeping every banked pairing comparable.
  • Groups: 8 seeds × 8 draws, temperature 1.0, advantage clip ±2.0 — unchanged surface-A shape.
  • Levers (draft proposal, finalization decision): lr 1e-6, kl_beta 1.0. Rationale to be judged at finalization: R1-B’s 3e-7 measured a greedy policy digit-identical across waves (it priced a null), and the escalation that motivated the ÷3.3 re-price was the shove-funding failure mode — since retired by v2
    • the competence change. 3e-7 stays the registered fallback if wave telemetry runs hot. kl_stop 0.06 unchanged.
  • Tripwires: the full R1 §7 set inherited verbatim — strikes, non-finite loss, spread collapse ×3, knock-away 2× ×3 (the in-reward fix does not retire the belt), kl_stop, competence-floor CI. setback_frac now HAS a baseline (R1-B banked 0.703 / 0.5625 / 0.5938); promoting it from record-only to a wire is a finalization decision, default record-only.

§3 Registered reads

  • PRIMARY: held-out sim100 (frozen seeds 0–99, harness success_tick) at the boundary vs the banked stage-D base count — paired per-seed, exact test. Success count materially above base with the paired read clean → accumulation confirmed, the SFT+GRPO alternation flywheel (success-recycle into the demo set) prices its next rung. Flat with wires quiet → banked negative: competence was necessary but not sufficient at this pressure — a real result, not a retry license.
  • Wave-0 calibration read (the success-variance claim’s first live contact): fraction of groups mixed (predicted ≳ 60% at p ≥ 0.2 base), advantage mass attributable to the ±10 term. Registered bar patterned on R1-B §4: if the groups do NOT carry success variance at the measured base rate, that is a calibration FAIL and stops the run before GPU-hours burn — the group shape (seeds × draws) is the amendment path, not lr.
  • Behavior reads (record + judge at boundary): knockaway_frac / setback_frac / earned-vs-ungrasped decomposition on a competent base — R1-B’s finding was that these are competence artifacts; the clean prediction is they FALL as success rises. If knock-away re-fires on a ≥ 20% base, the competence-artifact story is wrong somewhere and the wire analysis is the deliverable.

§4 Falsifiers

  • F-flat — success does not move at gentle pressure on a competent base: GRPO-on-this-surface banked negative twice (incompetent AND competent base); the 90% path re-prices toward more SFT rounds / demo scale instead of RL.
  • F-instability — wires fire early on the competent base: the pricing (lr/β) is wrong for this base, one registered re-price amendment allowed (the R1 ladder’s discipline).
  • F-regression — held-out success materially BELOW base at the boundary: KL anchoring insufficient; stop, bank, and the alternation flywheel inherits the diagnosis.

§5 Cost (measured pace)

R1-B measured ~0.98 GPU-h/step including the per-step held-out eval. Proposal: 10 steps ≈ ~10 GPU-h, gate ≤ 12, mid-ladder read at wave 5, ~30-min babysit cadence, self-stops armed. This is a fresh budget — the bootstrap’s ≤ 13 gate closes with stage D and does not carry over.

§6 Finalization checklist (fills when stage D banks GRPO GO)

  1. Stage-D verdict + the measured base success count (the PRIMARY’s anchor) recorded here.
  2. Base checkpoint id + conversion receipt (the exact bijou dir stage D evaluated).
  3. lr decision (1e-6 primary vs 3e-7 fallback) judged against the stage-D read + a dry advantage-decomposition pass on wave-0 rows.
  4. setback_frac: wire or record-only.
  5. Rollout seed policy for training groups (fresh stream per the standing variance branch; eval seeds 0–99 stay the frozen holdout, never trained on).
  6. HEAD re-pin, objection window opened in-channel, owner go or window expiry before launch.

§7 Amendment A1 (pre-finalization, 2026-08-15 10:3xZ) — the head seam, and two owner decisions

Registered mid-ride after the owner’s in-channel review (09:49–10:14Z exchange); nothing here was launched, so this amends a DRAFT.

The seam (owner-caught). As drafted, §2 is inconsistent: stage D measures the flow head (euler-10 through the converted action expert — the only module stage C trains), but the named serving stack (MolmoAct2DiscreteStack) trains the AR/FAST token head, whose weights stage-C SFT never touches (--ft_vlm=false). Token-GRPO from that base would start from the released model’s discrete policy — not the competent base this pre-reg’s §1 premise requires. The “byte-identical to the policy stage D measured” claim held for the checkpoint dir, not for the policy the loop would actually push on.

Owner decisions registered (2026-08-15, in-channel):

  1. Their trainer is retired for us (10:07Z): no further runs on molmoact2 train_lerobot.py; all training goes through bijou.train / first-class objects. The stage-C AR run (killed 10:11Z at step 2040 on the owner’s order, checkpoints 500–2000 retained) was the last their-stack run.
  2. Direction for this pre-reg: if token-GRPO is the tool, a token-SFT arm must precede it — a bijou.train run (first-class ar objective) over the same demo set, so the discrete head has measured competence before RL pressure. That arm would need its own pre-registration and its own sim eval; this R2 draft then re-bases on that endpoint (or re-scopes to whichever head the owner picks).

Status. The bootstrap’s stage-D formalism is suspended pending the owner’s re-steer after the step2000 probe (train-vs-unseen seeds, live at amendment time). §§1–6 above are retained verbatim as the registered record of the draft’s reasoning; activation now requires BOTH a competent-base verdict on the head GRPO would train AND the owner’s choice of route through decision 2.

§8 Amendment A2 (pre-finalization, 2026-08-15 14:4xZ) — the token-SFT arm is now a registered draft; re-basing spelled out

A1’s decision 2 said the token-SFT arm “would need its own pre-registration”; that document now exists — token-SFT arm pre-reg (DRAFT) — and this amendment binds the two:

  1. Re-basing. If the owner routes token-GRPO (route B or C in the arm’s §4), R2’s base policy (§2 bullet 2) becomes the token-SFT arm’s endpoint, and the activation bar becomes that arm’s primary read: unseen sim100 ≥ 20/100 on the discrete head under greedy decode — the same competence bar, measured on the head this run actually pushes. The stage-D flow-head verdict no longer activates this page (it measured the wrong head; A1’s finding). §3’s PRIMARY comparator becomes the token arm’s banked unseen count.
  2. Table lineage. The arm inits from the corrected-table base (verified in code this session: bijou/fast/codec.py normalizes token targets with the baked q01/q99), so R2 inherits the corrected lineage automatically — the §6.2 receipt must show it.
  3. Checkpoint receipt format. Owner main 4fd6875 landed the phase-3 VLA checkpoint format after this draft was written. §6.2’s “checkpoint id + conversion receipt” is now spelled: the pinned base carries a convert_legacy-produced (or natively written) VLA-format dir passing validate_checkpoint, whose metadata stats block + stats_note record the corrected-table provenance first-class.
  4. Unchanged. Reward v2/v1 split, 8×8 groups at T=1.0, tripwire belt, lr proposal + fallback, ≤ 12 gate — all stand. §5’s pace anchor (R1-B ~0.98 GPU-h/step) remains the estimate; it was measured on this same discrete stack.

§9 Amendment A3 (ACTIVE, 2026-08-19 13:2xZ) — activation from 7%: the band amendment

Status: ACTIVE — owner-activated 13:25:15Z 08-19 (“Re: my calls, I agree with all your recommendations”, message …407784, replying to the 11:19Z summary …522815 rec AMEND + ACTIVATE). §6-style finalization satisfied at activation: HEAD re-pin 570e53e (the commit carrying the A3.8 launch kit — flags, verdict instrument, staged launcher — and this page’s A3 text, frozen 12:5xZ before the reply); objection window satisfied by the frozen page + the in-channel summary + the owner reply. Execution per A3.8: preflight leg 0 fires at activation (H100 idle, policy-server down, verified), the A3.4 run fires on a PASS verdict. Drafted as an owner call under the pre-delegation register; superseding activations fall under the 2026-08-18 no-GO-asks delegation.

Plain words. The robot’s token-based control passed only 7 of 100 test scenarios — below the 20 we said reinforcement learning needs. But the autopsy of those failures showed something specific: the policy isn’t confused, it’s timid. It reaches the boat, closes the gripper, and then moves it at less than half the speed its flow-based sibling manages, so time runs out. That timidity is a known artifact of always picking the single most likely action (“greedy” decoding). Reinforcement learning as we run it doesn’t use greedy decoding — it samples, which is precisely the knob that relieves timidity — and its reward directly pays for finishing. So we propose to amend the entry bar and run RL from the 7% policy, with two cheap pre-checks that abort the run early if this reasoning is wrong.

A3.1 What the band said, what was measured

A2.1’s activation bar: unseen sim100 ≥ 20/100 greedy on the head GRPO trains. Measured (route-C joint step_002000, token head, probe of 08-18/19): 7/100 — band 5–19 → owner decision. Frozen anchors for every read below: token greedy 7/100 (success seeds 34, 35, 63, 68, 71, 91, 96), flow sibling 44/100 (same trunk, same scenes), token_base 0/100 (released model, SFT delta +7).

A3.2 The case for amending (the [decode

diagnosis](https://mcobzarenco-fontaine-reports.static.hf.space/eval__grasp_sft_joint_step2000__token_unseen100.html), frozen in analysis__token_decode_diagnosis.json)

  1. Not decode collapse: 0/300 frozen episodes across all three legs (motion instrument; min mean motion 1.41), no stereotypy. The head produces coherent, scene-directed trajectories.
  2. Greedy magnitude attenuation: funnel touch→pinch→success 60→22→7 vs flow 91→59→44 on the same trunk (base 7→0→0); conversion losses at both hops (touch→pinch 0.37 vs 0.65, pinch→success 0.32 vs 0.75); carry speed 0.81 vs 2.00 cm/s (ratio 0.405), 2 timeouts still holding the boat (flow 0); reach envelope truncated — 1/14 touch at 11–13 cm spawn, zero successes past 10 cm (flow: 17 there). The failures are under-magnitude versions of the right behavior.
  3. Iterate-once re-buys the wrong thing: CE owned all 2000 route-C trunk updates (joint objective, λ=1.0) — the token head is not undertrained by omission. Another token-SFT pass extends reach; it is not a targeted fix for greedy-mean attenuation.
  4. R2’s pressure lands exactly on the failure: group rollouts sample at T=1.0 (greedy never appears in training), and the +10 success term pays directly for completing the carry the greedy read leaves unfinished. Context, recorded as a rhyme and not evidence (different trunk, different metric): the #19 dT record measured AR competence monotone in T (6.50/6.57/6.78/7.18 at T=0.5/0.7/1.0/1.3, er60k).
  5. The §1 arithmetic at p = 0.07: mixed-group probability 1−(1−p)⁸−p⁸ ≈ 44% of groups carry a ±10 contrast — half the ~83% at the gate minimum, but ~4 in 10 groups informative vs ~34% at R1-B’s measured 0.05 base, and the sampled rate should sit above the greedy 7% if point 2 is right. The signal is plausibly there; the two gates below make that a measurement instead of a hope.

A3.3 Registered gates that replace the ≥ 20 bar

  • Preflight leg 0 (F-premise, ~1.3 GPU-h): single-draw sampled T=1.0 sim100 (seeds 0–99) on the pinned base, before any training. Read: sampled count vs greedy 7. Materially BELOW → the relief premise is wrong → abort the run, route to iterate-once, bank the read. At-or-above → premise holds first-contact; the sampled count becomes the recorded training-decode competence floor.
  • Wave-0 calibration bar (from §3, re-priced): fraction of groups mixed, predicted ≈ 44% at the greedy floor. Mixed < 20% → abort before GPU-hours burn; the group shape is the amendment path, not lr.

A3.4 Frozen run spec (deltas vs §2 + A2 only)

  • Base (pinned): checkpoints/finetune/fontaine_grasp_sft_joint_corrected/step_002000_v2 — schema-2 VLA dir, family joint, corrected-table stats first-class in metadata.json (A2.2/A2.3 satisfied); HF-recoverable bitwise per the 08-19 evacuation audit. Load seam verified in code this session: MolmoAct2DiscreteStack.load → read_metadata + load_vla; the joint family carries the format-6 discrete decoder — no conversion step needed.
  • Recipe unchanged (§2 + A2.4): composite_reward_v2 trained-on / v1 held-out, 8 seeds × 8 draws T=1.0, advantage clip ±2.0, clip-higher [0.8, 1.28], lr 1e-6 primary / 3e-7 registered fallback, kl_beta 1.0 vs the frozen base anchor, kl_stop 0.06, option-B trainable surface (text stack fp32, vision frozen bf16).
  • Seeds (re-pinned): --train-seed-base 2000 — the loop default 1000 now collides with the stage-B collection band 1000–1099 (demo corpus + measured train-probe band). Eval holdout 0–99 never trained on; in-loop 20-seed eval band 200–219 stays the loop default, record-only trend.
  • Knockaway wire (re-pinned): the config default baseline 10/120 ≈ 0.083 is an R0-A-era er60k measurement; this base’s greedy unseen read measured 25/100 knock-aways, so the old 2× line would fire spuriously at wave 0. Re-pin: baseline = the wave-0 measured knockaway_frac, wire at 2× ×3 unchanged. Instrument delta: expose --knockaway-baseline (config field exists; flag does not). setback_frac stays record-only.
  • NEW registered boundary read — flow-head regression leg (~1.3 GPU-h): the trunk is shared and option B trains the text stack, so GRPO moves the flow head too. Flow unseen100 (euler-10) at the boundary vs the 44 anchor, record + judge: material regression is F-regression evidence even if the token side improves — the joint checkpoint’s other head is not free.
  • PRIMARY (per §3, comparator re-based): boundary greedy sim100 vs 7/100, paired per-seed exact test (greedy stays the serving convention; RL pressure is sampled, the claim lands on the serving read). Record-only sibling: boundary sampled T=1.0 sim100 vs preflight leg 0 (prices decode-gap movement).

A3.5 Budget (re-priced) and kill rules

Pace anchor ~0.98 GPU-h/step (R1-B, same stack class, per-step 20-seed eval included). 10 steps ≈ ~10 GPU-h + preflight leg 0 ~1.3 + boundary token sim100 ~1.3 + flow regression leg ~1.3 ≈ ~14 expected, gate ≤ 15 (supersedes §5’s ≤ 12, which priced no boundary legs). Mid-ladder read at wave 5, ~30-min babysit cadence, detached unit + registry entry at launch. Kill rules: preflight F-premise abort, wave-0 mixed < 20% abort, the full §2 tripwire belt (knockaway wire re-pinned per A3.4), kl_stop 0.06, exit-3 self-stops armed.

A3.6 Falsifiers (unchanged + one new)

F-flat, F-instability, F-regression stand as §4 wrote them — F-regression now explicitly covers the flow-head leg. NEW F-premise: preflight sampled materially below greedy 7 — the attenuation-relief story is wrong, run aborts unstarted, the read banks against the diagnosis.

A3.7 Activation interface

Owner reply 2 ACTIVATE → same-session: §6 finalization checklist (HEAD re-pin, receipts, objection window noted as satisfied by this frozen page + the in-channel summary), launcher staged, launch at the next free boundary (H100 idle at draft time). 2 HOLD-BAND → token-SFT iterate-once arm per the registered band, this page holds. 2 PARK → the lane parks with the read banked.

A3.8 Launch kit (LANDED 13:xxZ 08-19 — code staged, still not active)

Everything above is now one command per stage (fontaine/scripts/launch_grpo_r2.sh, oracle-tested, full-parse green; the A3.4 argv is spelled exactly once):

  • parse-check (CPU): the frozen argv through sim.grpo_loop.parse_args with every A3.4 re-pin asserted, plus the babysit registry entry template (round-trip-verified against the babysit schema, A3.5 gates as gpu_hours_max 15 / vram_max_gib 75).
  • preflight (leg 0, ~1.3 GPU-h, detached unit): the sampled T=1.0 sim100 through the SAME probe driver and substrate pin as the greedy 7 leg (--serve-head ar --ar-temperature 1.0, --clutter-appearance standins), chaining into a machine-readable F-premise verdict (grpo_r2_preflight_verdict.py, provenance guards loud). “Materially below” is now pinned: exact one-sided binomial tail under Bin(100, 7/100) at the house 5% level — ABORT at ≤ 2/100 (P(X≤2) ≈ 0.026), PASS at ≥ 7 (the count becomes the recorded training-decode floor), BAND at 3–6 (below the anchor, not materially — a decision post owns it; the launcher refuses BAND without an explicit recorded override, and refuses ABORT unconditionally).
  • launch: refuses without a PASS verdict, then fires the A3.4 argv via run_detached.sh. The two in-loop gates are now first-class loop flags (oracle-tested, defaults unchanged — zero behavior change for every other run): --wave0-mixed-abort 0.20 stops the run at wave 0 below the A3.3 mixed-fraction bar (mixed_groups_frac is a new heartbeat key — success/failure contrast per group, distinct from the z-filter’s groups_kept), and --knockaway-baseline wave0 re-bases the violence wire at wave 0’s measured rate (the capture wave itself exempt), retiring the R0-era 10/120 pin for this run per A3.4.

A4 Wave-0 substrate postmortem + relaunch (2026-08-19 17:4xZ, decided + announced per the standing no-GO-ask rule)

What happened. The 16:10:02Z launch aborted itself at wave 0 (17:19:34Z, ~1.2 GPU-h): mixed_groups_frac 0.0000 < 0.20 (the A3.3 in-loop gate). The telemetry was not “unlucky scenes” — 64/64 training episodes (seeds 2000–2007 × 8 draws, T=1.0) showed zero successes, zero knock-aways, zero setbacks, ungrasped displacement ~0.0015, and the journal’s per-replan benchy->disk distances were bit-frozen from replan 0 to 29 on every seed. The step-0 in-loop eval read 0/20 on the same run. The boat was never touched, anywhere.

Diagnosis (convicted, two independent legs). (1) Code: the 08-18 clutter-patch promotion made patched the production default on the parallel driver’s WorkerConfig; sim.grpo_loop builds WorkerConfig without setting it, so the training waves AND the in-loop eval rendered patched clutter — while the A3.1 anchors (greedy 7/100, flow 44/100), the A3.2 decode diagnosis, and the preflight PASS (8/100) all ran the registered standins substrate, and the policy’s demo corpus is stand-ins-era. The promotion sweep pinned the eval drivers; the loop’s embedded rollout path was missed. (2) Measurement: the probe driver on the SAME seeds 2000–2007, same checkpoint, same T=1.0, standins — 6/8 episodes interact (progress up to 5.9 cm, min distance 3.9 cm, knock-aways present; 0/8 successes is unremarkable at p≈0.08, P=0.51). Substrate convicted; the 2000+ seed band exonerated (outputs/sim/grpo_r2/wave0_diag/seeds2000_standins_t1.json; the aborted wave banked at outputs/sim/grpo_r2/loop_wave0abort_patched/).

Reading. The wave-0 gate read an instrument artifact (substrate mismatch), not the registered calibration question (does success variance appear in groups at the measured base rate?). The A3.3 calibration claim is therefore still UNREAD — the gate stays armed at 0.20 on the relaunch. The gate itself performed exactly as registered: it stopped a doomed 10-step run after ~1.2 GPU-h.

Fix (mechanical, oracle-tested). --clutter-appearance {patched,standins} exposed on sim.grpo_loop (default patched — zero behavior change for any other run), threaded through BOTH the training-wave and in-loop-eval run_units seams, recorded in meta.json + the per-wave rows meta; the launcher pins --clutter-appearance standins in the frozen argv and parse-check asserts it. Oracles: parse-time typo refusal + a seam test that both wave closures forward the substrate.

Relaunch (same A3.4 run, registered conditions restored). The preflight PASS stands — it was measured ON standins, which is what the loop now renders; no re-run needed. Same frozen argv otherwise; all gates re-arm fresh (wave-0 mixed 0.20, knockaway wave-0 self-capture, kl_stop 0.06, tripwire belt).

Budget re-price (supersedes A3.5’s ≤ 15). Spent so far: preflight 2.25 (ran ~1.7× the greedy leg’s pace) + aborted wave ~1.2

  • diagnosis probe ~0.24 ≈ 3.7 GPU-h. Expected remaining: 10 steps ~10 + boundary legs re-priced at measured paces (greedy ~1.3, sampled ~2.25, flow ~1.3) ≈ ~14.9. Total expected ~18.5, gate ≤ 20. Kill rules unchanged.

A5 Serving-parity postmortem + launch gate (2026-08-19 19:4xZ — the fix behind the 18:06:48Z kill)

What the A4 relaunch showed (kill post 18:08Z). The 17:46:56Z relaunch — substrate genuinely standins, verified in meta and at the worker seam — read 0/20 at the step-0 in-loop eval with all 20 scenes bit-frozen replan 0→29, while the greedy anchor leg displaces the boat in 59/100 episodes (P(0 of 20 move) ≈ 2×10⁻⁸). The loop’s serving path is inert on the v2 base independent of substrate. Killed at 18:06:48Z, ~0.33 GPU-h in (lane total ~4.0).

Root cause (convicted in code, quantified). The retirement phase-4 re-point swapped the loop’s port predictor for MolmoAct2DiscreteStack but carried the port era’s hardcoded v30→v21 joint-frame shim (shoulder_lift mirrored, lift/elbow +90°) over unconditionally. That shim is correct exactly for a checkpoint whose recorded q01/q99 table is in the pre-PR#777 v2.1 frame — the unremapped official releases the port served (R1-B, which interacted, ran allenai/MolmoAct2-SO100_101, a v2.1-table HF layout). Every bijou-format table in play today is v3.0-frame (bijou-trained like the v2 base, corrected-table recomputes, and the conversion-remapped release alike — docs/so101-joint-conventions.md §3–6). On the v2 base the numbers are decisive: the shim maps sim lift ∈ [−103, +29]° to [61, 193]° against a table row of [−110, +12]° — the normalized value clamps at +1.0 on every frame (elbow likewise), so the policy is state-blind on the two joints that matter; and the decoded chunk comes back through the shim’s inverse (lift 90−a, elbow a−90), commanding poses outside the trained range every replan. Arm pinned off-workspace, boat never touched, distances bit-frozen — the exact telemetry of both kills. The doc’s §6 warned nothing automatic catches a missing/double remap; this was a missing-identity (extra) remap at a seam with no check.

Exonerated: the _batch seam flagged in the kill post (ACTION quantiles passed as state_stats) is inert — predict_ar on the ar/joint families detokenizes under the family’s own action_quantiles table and never reads the batch stats; prompt state normalizes in prompt_inputs before _batch is built. Left as-is.

Fix (landed this session, oracle-tested). --joint-frame {auto, rig, v30-to-v21} on BOTH sim discrete drivers (sim.grpo_loop, sim.rollout_sim_parallel --molmoact2-discrete), resolved through one helper (resolve_joint_frame) that reuses JointFrameTransform’s literals (the rollout CLI’s vocabulary — rig = identity, exactly what BijouPolicy’s path applies): auto fingerprints the checkpoint’s state table per the conventions doc §4 (descending lift pair or far-negative lift → v3.0; lift ≥ +30 with elbow corroboration → v2.1) and refuses an unclassifiable table; an explicit mode that contradicts a classified table is refused too — the mismatch class is now unrepresentable at this seam, for old lineages and new. Resolved frame recorded in meta.json, the rows meta (state_units), and the driver out-json.

Parity oracle (CPU, in the suite). tests/test_joint_frame_parity.py: the classifier pinned on the three real table shapes; refusal semantics pinned; the shim literals pinned against JointFrameTransform both directions; and prompt parity — the loop stack’s prompt_inputs vs the BijouPolicy collator on the same observation collate to bit-equal MolmoAct2Inputs tensors on the tiny fixture. Beyond that seam the two paths share predict_ar and the family’s one quantile table, so prompt parity + the identity frame map IS serving parity.

GPU parity read (REQUIRED LAUNCH GATE, wired). launch_grpo_r2.sh parity (~0.7 GPU-h, next free GPU window — onerig owns the H100 to ~07:0xZ 08-20): seeds 200–219 greedy on standins through BOTH paths — the loop stack under --joint-frame rig vs BijouPolicy --serve-head ar. grpo_r2_parity_verdict.py rule, registered here: interacted := min_cm < initial_cm − 1e-6 or final_cm ≠ initial_cm; PASS iff |Δsuccesses| ≤ 2 AND |Δinteracted_frac| ≤ 0.30 (the convicted mode reads 0.00 vs ~0.59 — decades outside; the band absorbs decode-stack noise, which moves seeds, not fractions). launch now refuses without a PASS verdict — no override exists. The frozen argv gains --joint-frame rig (parse-check asserts it); all other pins and the A4 budget stand: spent ~4.0, parity +0.7, relaunch ~14.9 → expected ~19.6 against the ≤ 20 gate — no re-price, but zero slack: any further abort ends the lane at the gate.

§10 LANE CLOSE-OUT (2026-08-20 14:1xZ) — died at the gate; what the ~6.6 GPU-h bought

Outcome: NO primary read. The A3.4 relaunch (11:59:20Z on the perfect-parity PASS) died 13:59:42Z in the step-2 backward: torch.OutOfMemoryError wanting +1.47 GiB at 78.26 GiB in use. Step-1’s recorded peak was 72.09 vs the ≤75 gate — the step-2 update crossed the gate in-flight, caught by the allocator instead of a poll. save_every=5, death at step 2 → no checkpoint, resume impossible. Retry arithmetic: ~4.7 spent + ~1.9 burned + ~14.9 fresh ≈ 21.5 vs the A4 ≤20 gate; the registered zero-slack rule (“any further abort ends the lane at the gate”) applied verbatim. Lane CLOSED, no step-10 sim100 read, no override.

Banked positives (what any R3 pre-reg inherits for free):

  • Serving parity PASS, perfect (~11:58Z): both paths 2/20, same seeds 207/214, interacted 20/20 both, Δ 0 / 0.0 vs tolerances 2 / 0.30. The convicted 0.00-interacted serving mode is dead; the --joint-frame rig classifier + parity oracle are permanent infrastructure.
  • Wave-0 calibration read: mixed_groups_frac 0.50 vs the <0.20 abort bar (predicted ~0.44) — the knockaway re-base works; the 7%-base premise (§9 A3) survives first contact.
  • Step-1 row healthy: kl 0.0041, strikes 0, groups 8/8 kept, 6/64 wave-0 successes, knockaway_frac 0.3281 banked as the violence-wire baseline.

Death mechanism + the R3 lead (registered, not priced): grpo_loop.py sets no PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, and episode-length variance swings backward memory ~6+ GiB step-to-step at a 72 GiB baseline. An R3, if ever re-priced, needs the alloc conf + measured headroom in its pre-reg — or a smaller group/batch geometry. No R3 is scheduled; the freed window went to the demos+clean poison-pinning cell (pre-reg 2026-08-20, delegated).

Lane ledger: preflight 2.25 + patched wave ~1.2 + probe 0.24 + relaunch-1 0.33 + parity 0.66 + A3.4 run ~1.9 ≈ ~6.6 GPU-h spent of the ≤20 A4 lane gate; death post 1539998329893556224.