Pre-registration DRAFT: rig-mixture fine-tune rung on the er_60k trunk

Drafted 2026-08-11 ~16:5xZ (work session), executing the mixture lever the er-60k pre-reg pinned and the ER screen close named as the next unpriced step. Status: DRAFT — HOLDING for the owner compute call (the 4× box is gone; see “Compute ask”). Finalizes into an immutable pre-reg with a param sheet in-channel before any GPU minute, per the standing gate.

Plain words

Our best trunk (er_60k) trained on ~18.7M frames of community robot data plus the owner’s two rig datasets — but the rig data is only 0.19% of the corpus, so the model effectively never saw it (each rig frame ~0.9× in expectation over the whole run). This rung continues training from the er_60k endpoint with the rig data oversampled to ~5% of every batch, and asks two questions with one run: does the model get meaningfully better on held-out rig episodes, and does it do so without getting worse on the general panel? That pair is the whole VLA-for-the-rig question in miniature: one model that works on the owner’s robot without forgetting everything else.

Instrument (landed, this session)

The loader dedups repeated roots, so there was no zero-code oversample. --dataset-repeat PATTERN=COUNT (commit 1b1c314) replicates matching datasets in the concatenated train set after all guards and the episode split: fnmatch against <user>/<dataset> repo ids, first matching spec wins, a spec matching no selected dataset is fatal (a silently unapplied oversample would corrupt the registered mixture). Same objects — no extra host RAM; the shuffle, length-bucket keys and DistributedSampler all see the replicas. Training-only: eval call sites never pass it. Oracle test tests/test_dataset_repeat.py pins parse/precedence/no-match plus the mixture arithmetic below; check.py green (683).

Design

One arm, warm-started from the reference trunk:

• Init: --init-from er_60k/step_060000 (weights-only is sufficient and is what survives the box teardown — fontaine-checkpoints, commit 4ed3dd0; the optimizer state is gone with the box, so a fresh optimizer is forced, not chosen).
• Mixture: --dataset-repeat mcobzarenco/so101_pick_place_clean=27 mcobzarenco/so101_pick_place_v2=27 — explicit per-repo specs (no wildcard: immune to future owner datasets landing in the curated collection). Effective share = 27×36,078 / (18.67M + 26×36,078) ≈ 4.97% (natural 0.19%), inside the CL-triangle 2–20% replay band (page).
• Steps: 10,000 at eff-batch 48 — ~24k rig draws ≈ 0.66 rig epochs. Named tension: the full-run arithmetic the mixture note imagined (~4 rig epochs) belongs to a 60k-step run; a screen-sized rung at 5% is sub-epoch. If the owner wants a rig-heavier rung instead, ×129 ≈ 20% share (top of the CL band) gives ~2.7 rig epochs in the same 10k steps — same cost, more forgetting risk. Default registered arm is 5%.
• Recipe: er-60k launcher verbatim otherwise (decoder ar_backbone, aux + condition fields, decoder-lr 1e-4, backbone-text-lr 2e-5, fps 30, camera-counts 1 2, holdout 0.1 split-seed 0, async saves). Named deltas: --warmup-steps 500 (fresh optimizer re-estimates moments; 1000 would be 10% of a 10k run), --save-every 2500, --seed 3 (fresh-seed-on-extension standing rule; 0 and 2 are used in this lineage), cosine decays over the rung’s own 10k steps (the init-from convention).
• No control arm registered (startup-velocity): at natural share a 10k continuation sees ~450 rig draws — it cannot explain a material rig gain, so the er_60k endpoint itself is the baseline. A natural-share control continuation is the named escalation if the primary read lands small or ambiguous.

Frozen reads (numbers finalize at param-sheet time)

1. Primary — rig holdout, paired CI95: the deterministic episode holdout (fraction 0.1, split-seed 0) holds out 1 of 7 clean + 5 of 50 v2 episodes (~3.7k frames) that er_60k never trained on. Score mixture endpoint vs er_60k/step_060000 on identical frames (+ state-copy anchor), paired per-frame CI95. Pass = mixture below the er_60k baseline with CI excluding zero. This is a genuine within-rig generalization read, unlike the contaminated-by-construction MolmoAct2 rig-ft train-frame reads.
2. Guard — panel non-regression: k4l2 panel_v2 eval at endpoint (--report + npz per the standing rule), paired per-frame vs the banked er_60k endpoint panel (5.7782/1.9898). Fail = pooled delta worse than +0.05 with CI excluding zero (band draft — finalize against the 55k→60k rung step of ~0.05).
3. Record-only: probe ladder at the er-60k cadence; aux-head accuracies vs the endpoint row (holding 0.915 / progress 0.060 / event 0.858 / visible 0.822); per-episode rig MAE spread.
4. Kill lines: NaN/inf; probe divergence bar re-derived from the er_60k tail at finalization; vram near-OOM per first-poll actuals.

Compute ask — OWNER DECISION, the run is not launchable without it

The 4× box is torn down. Options, priced:

• (A) New 4× box rung — recipe verbatim, ~2.5 GPU-h/1k steps measured on the er_60k run → ~28 GPU-h, ~7 h wall incl. the endpoint panel. Cleanest: zero recipe deltas beyond the named ones. Gate 32 GPU-h.
• (B) Local 1×H100 — the full recipe is measured structurally OOM single-GPU (08-08 perf review: 78.2/79.18 GiB by step 2, batch-invariant, ZeRO-1 unshards +11 GiB). A local leg therefore needs a fit-preflight ladder first: --activation-checkpointing (landed + oracle-gated, ~2.4–2.8 GiB/sample) + small batch × backward-chunks, ZeRO-1 off. If green: ~same GPU-h, ~14–16 h wall (serial 4× + recompute overhead), local GPU blocked for a day. If red, (A) is the only path.
• (C) Defer — the instrument and this draft keep; the rung launches whenever compute lands.

Non-goals / non-deltas

No vision unfreeze, no AdamC, no recipe retune, no rollout claims (rig rollouts are the runbook’s out-of-band path and need the convention rails). The AE-on-our-trunk work is a separate queued item behind the main rebase and does not ride this rung.