Molmo2 AR 40k endpoint: BEATS — the trunk bet pays

2026-08-08 05:0xZ. Reads frozen in the pre-registration §5, executed by fontaine/scripts/molmo2_endpoint_results.py (oracle-green before the data landed; output reports/analysis__molmo2_endpoint_k4l2.json). Run: fontaine_molmo2_ar_40k_ddp4, 40,000 steps on 4×H100 (~29 h + saves), Molmo2-4B trunk, aux + subgoal co-training, endpoint probe 6.2075@40000 (run low 5.91@26500). Panel: panel_curated_v0_k4l2, 25,800 frames / 17,204 core, chained greedy eval at the endpoint.

Read 1 (primary): BEATS, by 3.4× the bar

Classification bar: BEATS < 7.30 (anchor − 0.5, ~the family’s seed spread). Measured: 6.0079 — not a margin call. The paired per-frame read (identity-aligned rows, seeded frame bootstrap): Δ = −1.717 [CI95 −1.797, −1.635] over all 17,204 core frames. First-step MAE nearly halves (2.19 vs 3.94).

Frozen decision executes: Molmo2 becomes the phase-2 flow-trunk candidate — the AR-adapted prefix is what the attachment screen (#4, launching next on this box) will hold frozen, killing the −2.7 topology confound the pre-reg named.

Context the board gives for free: at 40k steps the Molmo2 trunk sits 0.21 behind AR-100k’s greedy 5.8026 at 2.5× fewer steps, on a trunk that reads our rig scenes zero-shot. The #19 draws10_t1 arm (running now, same chain) prices its sampled-ensemble headroom.

Read 2 (instrument integrity): green, with one recorded slip

The state-copy and state-copy-norm columns byte-match the banked same-plan panel columns — same rows, same fallbacks, no instrument drift. Pooled state-copy quotes 11.7847/2.6202 under the panel convention. Recorded, not silently corrected: the pre-reg’s parenthetical quoted “11.7639/2.5851”, which reproduces under no pooling of this plan (core, all-rows and norm variants all checked) — a drafting slip in the pre-reg text; the operative byte-match oracle is unaffected.

Read 3 (context, narrative only)

The e4b screen milestone family (7.54@10k probe) and arm C statedrop (10.50) both sit well above this endpoint. Nothing here changes those banked readings.

What was consumed, what it cost, what broke

~29 h train on 4×H100 (2.17–2.55 s/step; ~4 h of wall clock was the save windows) + ~0.5 h chained greedy panel eval. vram_alloc_peak 67.13 ≤ 71 GiB throughout; K1 kill line crossed green at 10k (7.1652 vs 12.0944) and never looked back.

One incident at the boundary: the chained greedy eval died 4 minutes in on its first launch — float != BFloat16 at the first suffix-attention matmul. The suffix decoder’s torch.where silently promoted mixed-dtype embeds (bf16 mounted trunk, fp32 trainable FAST patch); training probes run under autocast, so this standalone-eval path had never executed against a real bf16 checkpoint, and the tests loaded the tiny fixture in fp32. One-line cast fix (the idiom _logits already used), a red-then-green regression test (test_bf16_mounted_trunk_decodes_with_fp32_patch), and the #19 launcher’s pre-built greedy-if-missing clause recovered the chain (5a43b15). The dead-chain survival path being designed in advance is why this cost ~10 minutes, not a session.

Consumables

• Endpoint weights (backbone/expert/prompt + config, weights-only) uploaded to mcobzarenco/fontaine-checkpoints under fontaine_molmo2_ar_40k_ddp4/step_040000 — machine-loss protection per the standing rule; optimizer state stays box-local.
• For the #17 vu5k finalization amendment (execution cell 2): the frozen-sanity bar input is the endpoint probe 6.2075@40000.
• The greedy npz (eval__fontaine_molmo2_ar_40k_ddp4__step_040000__panel_curated_v0_k4l2.npz) is the paired baseline for the #19 draws arm’s Δ_AR read.

Leaderboard: molmo2 greedy row added (decode-cost cells pending the queued microbench — molmo2 configs were not in the measured set; nothing mtime-derived is quoted).