17. New trunks / new architectures — standing owner mandate (2026-08-05 17:24Z)

Tag: new-trunks · idea #17 · index

“The current repo should just be a starting point for what I’ve tried.” A ranked exploration front, fed by the literature slice; every candidate enters at the screen rung with a pre-reg and counts toward the exploration budget.

• Lit feed 2026-08-07 (page): trunk-screening criterion from VLM4VLA (2601.03309) — general VQA-bench scores are POOR predictors of VLA rank (Kosmos-2 1.7B beats 30B-class models on SimplerEnv; no model dominates across suites); the load-bearing component is the VISION pathway (frozen-encoder collapse 4.057→2.823 Calvin, worth +29 pts when action supervision reaches it). Screening a candidate trunk = probe its vision-pathway adaptability on OUR data, not its benchmark card. Also: all 7 of their embodied-VQA co-training mixes UNDERPERFORMED plain baselines — do not import aux-data recipes on faith (compare against our own #6 measurements).

• Vision-unfreeze rung pre-reg DRAFTED 2026-08-07, AMENDED to the warm-start two-arm design same day (owner steering 18:02Z) (draft post; owner question 17:04Z + the vision-encoder-freeze slice) — both arms --init-from the 40k endpoint checkpoint (frozen-continue control vs thawed-continue --backbone-vision-lr 2e-6; full-FT tower, never LoRA-on-SigLIP), 3k steps each at the 40k tail LRs (decoder 1e-5 / text 2e-6), seed 1 both arms → identical batches, fresh AdamW symmetric (--resume mechanically excluded: the extra vision param group breaks optimizer.load_state_dict). Primary = thawed@3k − frozen@3k paired per-frame Δ (CI95, null band 0.07; critical-frame re-pool robustness). ~15 GPU-h train vs ~27 from-scratch, and a win directly upgrades the deployment artifact. Declared caveats: late low-LR thaw may understate from-scratch unfreeze (tie ≠ “unfreezing doesn’t help”); panel can’t see the MAPS-style OOD tax. Memory ladder unchanged (67.07/71 + ~3–4 GiB tower adder → chunks 6→12 → decoder activation-ckpt; matched downshift excluded). DRAFT status: execution blocked on finalization amendment (launcher byte-audit + memory smoke + endpoint probe quote) + owner go, window post-attach-screen (~08-09+).

• Amendments 2+3 (owner exchange 18:3x–18:51Z 2026-08-07, “Ok, agreed”) + finalization PREP landed same day (485194b): 5k steps/arm (was 3k), gate 32 GPU-h; LRs = 0.3× reheat of the 40k peaks (decoder 3e-5 / text 6e-6, fresh 5k cosine to the 10% floors — pure tail LRs were judged a null-bias on exactly the co-adaptation axis), warmup 200→500, vision LR 6e-6 tied to the text group (was 2e-6). Prep item executed 19:4xZ: amendment-3 flag set byte-audited clean against bijou.train at HEAD, both arm launchers landed (launch_box_fontaine_molmo2_vu5k_{frozen,thawed}_ddp4.sh — arm-vs-arm diff is exactly the one flag; thawed refuses without the frozen endpoint AND the vu5k_mem_ready smoke record) + prepared babysit entries. Remaining before launch: 150-step thawed memory smoke from the endpoint checkpoint, endpoint-probe quote, amendment POST, owner go.

• OWNER PICK (2026-08-05 21:57Z): E4B screen confirmed as the next pre-reg — AR-100k on the freed 4×H100, matched parameters with the E2B AR-100k run (verified from the recipe: --batch-size 12 /GPU on DDP4 = effective 48, decoder-lr 1e-4, backbone-text-lr 2e-5, grad-clip 100, seed 0, same aux/condition/dropout flags; if E4B OOMs at 12/GPU, grad-accumulate to the same effective 48 — stated in the pre-reg, batch semantics never change mid-comparison). Gates = the MAE curve over time, not the endpoint: matched-cadence in-run probes vs the banked E2B curve

  • mid-run panel evals (~25k/50k) with pre-registered bands so a losing rung dies early. Owner 21:58Z: image (embedding) budget is the follow-on ablation arm on the winning trunk — one variable per rung, trunk first, then image token budget (pairs with #11’s grounding read). Pre-reg POSTED 2026-08-05 ~22:4xZ (post): verbatim mainline recipe + --backbone google/gemma-4-e4b-it (AR path verified fully config-driven), eff-48 with a pre-registered chunked-backward fallback ladder (bijou.train has no grad-accum today — impl + oracles is a pre-launch CPU item if the memory smoke says B12 doesn’t fit; E2B B12 peaked 77.5 GiB, E4B text ~2.2× params), probe-curve gates @30k/50k (±0.5 floor) + mid-run panels @25k/50k on the local GPU, endpoint adopt rule bound to tonight’s E5 σ_seed via finalization amendment. Launch blocked on: box free + e4b checkpoint download (not in box cache) + parity spot-check + memory smoke + amendment.

• NO-LAUNCH (2026-08-06 ~05:4xZ) — the pre-registered memory ladder exhausted: all four rungs (B12 direct / 2×6 / 3×4 / 4×3 chunked backward) OOM’d on 80 GB before completing one optimizer step; Adam state (~31.8 GiB) never even allocated ⇒ steady-state needs ~≥110 GiB/rank under the matched recipe (finding post, Amendment 2 of the pre-reg). Feasibility negative, not a scale answer — the probe/panel gates never ran; the attribution question stays open. Follow-on = owner decision (options posted): ZeRO-1 re-entry as a NEW pre-reg vs redirect to Molmo2-4B (rank 2) / #11 grounding arms. E4B’s zero-port-cost premise is dead; Molmo2-4B competes on closer-to-even terms now.

• Owner steering 2026-08-06 11:44Z — multi-GPU architecture run requested (new trunk / full residuals / bigger images, “really just examples”) → architecture batch #1 pre-registered ~12:2xZ (pre-reg, filed under #11 — the owner’s examples are the grounding front): bigger-images

  • full-residual arms first (same trunk, clean attribution); the trunk swap (Molmo2-4B) is its own follow-on pre-reg, promoted to next-in-line if both arms null. E4B ZeRO-1 re-entry queued behind the architecture run (owner 11:44Z: E4B paused).

• Arm B implementation LANDED 2026-08-06 ~12:4xZ (the F1 critical path: the smoke needs BOTH configs, so arm A could not launch before this existed): --conditioning-streams residual in bijou.train — encoder exports raw post-layer hidden states (res0..res14), the flow expert projects them through learned per-layer adapters (RMSNorm + K/V proj + k_norm/v_norm + RoPE, mirroring TextAttention.project_kv exactly, so the streams are contract-identical to K/V exports and the blocks are untouched). Adapters live DECODER-side and attach OUTSIDE the no-grad prefix encode — trainable under the frozen trunk. Real-config count 23.62M params ✓ (pre-reg said ≈23.6M). All five pre-launch oracles green as CPU tests (tests/test_residual_streams.py, 11 tests): stream contract + padding-orientation invariance, trunk bitwise-frozen through an optimizer step, grads reach all adapters, checkpoint round-trip with no flags, K/V path untouched (state-dict keys + banked loss oracles). check.py 285 green. Still owed at the arm-C-boundary code sync: SnapFlow stage-0 re-verify on the box + F1 two-config smoke.

• Molmo2-4B port plan POSTED 2026-08-06 ~14:1xZ (owner-promoted 12:03Z, background work independent of the batch verdict; plan, distilled primary- source doc docs/molmo2.md per §6 post-cutoff rule). Key design calls: residual-only conditioning (arm B’s path — learned adapters keep the expert contract at kv1×512 regardless of Qwen3’s GQA 32:8; no KVCache/layer-type/project_kv port needed) and 15-of-36-layer mount (fractional depth 0.417 vs E2B’s 15/35 = 0.429 — expert depth and res0..res14 schedule carry over unchanged). Five WPs: WP0 seam refactor (the docs/plan.md encoder ABC, oracle-guarded, lands alone) → WP1 Qwen3 decoder port + HF parity (shared with InternVL3.5/Qwen3-VL — one port, three trunks) → WP2 SigLIP tower/connector → WP3 ChatML collator (turn-close probe + state-slot splice re-proved) → WP4–5 exports/schema/audit. Phase 1 = flow on the raw frozen prefix (no AR port, no vocab surgery; the AR-adaptation −2.7 confound ships with any claim). Mounted footprint ~2.3B ≈ 4.7 GiB bf16; est. 4–6 CPU work sessions, GPU only for parity bursts + the memory smoke. First run gets its own pre-reg after the §4 oracle suite is green.

• External prior (lit slice 2026-08-05 ~22:5xZ): 2606.31382 (VLM-to-VLA parameter redundancy) reports bigger VLM backbones do NOT consistently improve action performance after adaptation (their ablations; skim-depth read via a fast-model summary — re-read before citing numbers). Direction: strengthens the E4B screen’s kill branch as a live outcome, not a formality — and raises the prior on #11 (grounding/adaptation quality, not trunk scale, as the binding limit). The screen runs regardless: our recipe, our corpus, pre-registered either way. Papers-page re-read 2026-08-07 (page) — banked claim CORRECTED: 2606.31382 makes no backbone-scale claim at all (it is a pruning-as-diagnostic study: divergence-ranked, no-recovery pruning removes 12–30% of params at 85–96% retention); the bigger-isn’t-better claim belongs to VLM4VLA, which it merely cites and which we already carry via the ICLR-26 survey — cite VLM4VLA for the kill-branch prior, not this paper.

• Lit slice 2026-08-07 ~04:0xZ — world-action models on the radar (via the NVIDIA WAM post; skim-depth, re-read before citing numbers): the emerging tier above VLA trunks conditions action decoding on video-model dynamics — UniPi ’23 → GR-1 ’24 → DreamZero ’26 (monolithic video+action denoising; RoboArena 1750 vs π0.5’s 1622), LingBot-VA (Wan 2.2-5B inverse dynamics), Being-H0.7 (latent VLA↔WAM bridge); π0.7 itself now renders subgoal IMAGES from a BAGEL world model between the HL policy and the action expert. Direction: a video-capable trunk (Molmo2 — already ours) plus a subgoal-image conditioning arm is the reachable-scale version of this thesis; ties the #6 explicit-HL ladder to the trunk front. No action until the Molmo2 40k endpoint + stage-2 decision land. Papers-page re-read 2026-08-07 (page): all named systems + numbers verified; two additions — Fast-WAM (representation-only, skips test-time video generation, 3–4× faster, reportedly matches LingBot-VA without the 16k-h robot pretrain) is the strongest evidence the video prior not the generation carries the value → the reachable Molmo2 version is predictive-feature conditioning, not rendered frames; and a #6 flag — π0.7 found TEXT subtasks insufficient for its bias-breaking tasks (needed rendered subgoal images), so a null on our rung-(a) text probe is consistent with the field, not fatal to the hierarchy thesis (state this in the pre-registered read).

Ranked 2026-08-05 by the trunk survey (paper + fetched-config deep-reads, owner method): 1. Gemma 4 E4B (zero-cost in-family rung) → 2. Molmo2-4B (best-in-tier quality, video+spatio-temporal grounding, Apache) → 3. InternVL3.5-4B (same Qwen3-4B decoder as Molmo2 — one port serves both; only modern 4B with a true base ckpt ⇒ the idea #10 vehicle) → 4. V-JEPA 2.1 ViT-L augmentation arm (the dynamics bet; 2-AC’s <62h-robot-video → zero-shot Franka is the strongest external evidence for the north-star thesis) → 5. Qwen3-VL-4B in reserve (most reimplementation surface, no base ckpt). Screened out: Ministral 3 3B (no video), SmolVLM2 (older gen; but SmolVLA = external validation of our trunk+flow-expert protocol), Cosmos-Reason1-7B (second round iff E4B says scale is the lever), all MoEs (budget + export-stream semantics). Original slate below, kept for scope:

• Trunk swaps at reachable scale: E2B → stronger open VLM families (Qwen-VL, larger Gemma-4 variants — E4B/12B) through the existing stage-2 trunk-swap protocol; also the queued base-vs-IT swap (idea #10) as the cheapest member of this family.
• Video/dynamics-pretrained encoders (V-JEPA-style) vs image-language pretraining — the grounding probes (idea #11) say the visual stack is the bottleneck; dynamics-predictive pretraining is the structurally different bet.
• Tokenizer-free continuous action heads end-to-end — remove FAST; the flow expert reads the trunk at full depth rather than export streams {4,9,14} (subsumes the idea #4 stream question).
• Small world-model / latent-dynamics trunk trained on the community corpus, policy as readout — cross-embodiment by construction; the speculative end.
• Consistency-distilled 1–2-step deployment decoders (pairs with ideas #1 and #12) — the deployment-latency leg of the rig goal.

Lit 2026-08-08 ~16:2xZ (conditioning shortcuts, DISC 2605.20856): one banked prior for future conditioning-path debates (attach-screen seam variants included): conditioning delivered through a separate structural path grounds better than conditioning mixed into shared tokens — DISC’s combinatorial identical-scene bench (86.4% vs Octo 78.5%) is the cleanest number. Not an arm; its hypernetwork form costs fine-placement precision, disqualifying for a manipulation trunk.

• Field-level grounding evidence 2026-08-09 (fields panel, record-only): on the narrated-field table, Molmo2@60k ≈ AR-100k on the action-adjacent fields (event 0.880/0.878, progress at the label floor) but visible slot-set accuracy 0.819 vs 0.319 — the pointing-supervised trunk’s advantage shows up exactly where scene grounding is measured, first direct field-level support for the vision-side half of the Molmo2 bet (and consistent with VLM4VLA’s vision-pathway criterion above).

• Lit 2026-08-09 (VLAFlow): the V-JEPA-style bet on this page’s slate gets its first measured form — frozen V-JEPA-2 latents as auxiliary future-prediction targets (not as the trunk) were the biggest single transfer lever in a controlled 4-recipe bake-off. Cheaper entry point than a trunk swap; hook shared with #6.

• Lit 2026-08-09 (radar hooks cleared: VEGA 2605.10485 + HyperVLA 2510.04898): VEGA lands the third pole on the vision-freeze axis — an encoder-output alignment aux to a 3D-aware teacher (DINOv2-FiT3D), projector discarded at inference, beats LLM-token-level alignment (Spatial Forcing) on RoboTwin easy AND hard; its frozen-FiT3D ≈ unfrozen-FiT3D probe says unfreezing pays only while features lack what control needs ⇒ banked as the vu5k readout’s interpretation lever + the named cheap escalation if thawed wins (caveat: Molmo2 is single-tower — no clean “spatial branch only” split; teacher fragility real, VGGT-as-teacher collapses to 0.04 hard). HyperVLA stakes the inference-efficiency pole: understand-once/execute-tiny (0.1M generated policy per episode over shared DINOv2, 4 ms/step, 90× fewer activated params vs OpenVLA, sim-only) — trunk-ledger entry for the rig-latency conversation + the generated-update √d normalization rule (OOD-specific failure) for any future weight-modulation adapters; its MSE-beats-diffusion ablation is regime-bound (per-task specialist policies) and does NOT read onto AR-vs-flow.

• Lit 2026-08-09 later session (Spatial Forcing page 2510.12276, the banked VEGA baseline examined): the aux-alignment third pole now has TWO recipes with a measured teacher×depth interaction — SF aligns visual tokens at LLM layer 24 (of 32; “deep but not deepest” ablated: 24 ≫ 32/16/1) to VGGT by cosine via a discarded projector, and VGGT works there while VEGA saw it collapse to 0.04 at the encoder output. Teacher choice is not separable from alignment depth. SF’s headline is NOT final score (LIBERO ~parity, tables setting-ambiguous) but convergence: same success in ~50k vs ~150k iterations (“up to 3.8×”) + 5%-data +25.8 pp — a fewer-steps-to-quality lever, distinct from every step-time lever in the perf thread; teacher-forward training overhead unreported (demand it before any pre-reg). For vu5k: SF never tests unfreezing, so VEGA’s frozen≈unfrozen probe stays the freeze-axis evidence; escalation order stays VEGA-first (beat SF head-to-head on RoboTwin), with SF’s LLM-interior hook as the named sibling that may fit our single-tower Molmo2 better (no encoder/LLM seam needed).

• Lit 2026-08-09 12:1xZ (QDepth-VLA page 2510.14836, last banked radar hook cleared): the aux-spatial menu gains its third recipe class — {encoder-align (VEGA), LLM-interior-align (SF), expert-generative (QDepth: parallel 18-layer expert predicts VQ depth tokens from vision tokens, monocular pseudo-labels)}. The third is the only one needing NO encoder seam — vision tokens in, depth codes out — so it is the named single-tower-Molmo2 fallback if the family is ever pre-registered. Two teeth pulled before citing: the depth tokens ride the inference context (deploy cost nonzero, unmeasured — the family’s zero-cost selling point is traded away with no head-to-head vs VEGA/SF anywhere), and the ablation splits the +8.5 into ~−2.9 supervision / ~−5.6 scaffold — the aux-signal claim is the small half.

• Lit 2026-08-09 12:xZ — the trunk-redundancy ledger opens with numbers (Fewer layers / CLP page 2606.20246, deep-read same session the sweep banked it): CKA twin-layer pruning before finetuning (one calibration forward pass, keep the first of each high-similarity run, finetune heals the seams): 33–50% of π₀/GR00T-N1.5 depth removable — including 8 of GR00T’s 16 DiT expert layers — at ~−28-31% train time and ~−28-30% inference; low-data finetunes gain (+6.9 π₀ LIBERO at 10% demos = implicit regularization; full-data GR00T −0.9 ≈ cost-neutral, which is our regime’s honest expectation). CKA beats MSE/cosine/random selection. Banked: the CKA map as a one-forward-pass diagnostic for our trunk+expert checkpoints (the fractional-depth mount discussion has no redundancy evidence behind it); expert-sizing datapoint beside HyperVLA; throughput fourth lever class (FLOP-count mechanism — immune to the kernel-scheduling artifacts that killed perf pass-1’s microbench transfer). Prune-then-attach = named sequel arm for any future attach screen, own pre-reg required.

• Lit 2026-08-09 12:3xZ — the early-fusion pole staked (Qwen-VLA page 2605.30280, read same session): first production VLA on a natively early-fused trunk (Qwen3.5-4B, ViT tokens interleaved in the text stream, gated-linear hybrid attention) + 1.15B single-stream DiT flow expert (joint self-attention over concatenated VLM states + noisy chunk, AdaLN). Strong everywhere, and the OOD headline — real-ALOHA 76.9 vs π₀.₅’s 41.5 — is the claim early fusion is supposed to buy; stack-vs-stack confound loud, no fusion-controlled ablation exists. Four-stage recipe: T2A (expert trains, trunk FROZEN) → joint CPT → SFT (VL weight 0.1 vs action 1.0) → narrow PPO. Trunk-ledger entry beside VLM4VLA; the frozen-first Stage I is filed on #4’s ledger.

• 2026-08-09 fresh sweep — the adamc_100k grad-norm watch gets its interpretive frame (weight-decay correction page, 2512.08217, AdamC’s direct successor): re-derives AdamC’s λ ∝ γ from steady-state assumptions while refuting the orthogonality mechanism both AdamC and the rival λ ∝ γ² camp lean on (renormalized-AdamW control: deleting the perpendicular component changes ViT-S/16 top-1 by 0.3). Reads onto the live run: (1) head-exclusion partition validated twice over (Defazio’s own Llama-3 setup + this derivation — ours already excludes the head with a tied-param guard); (2) expected endpoint signature = flat grad/weight norms with ~nil final-loss effect (AdamC 76.98 vs AdamW 76.92), matching the record-only framing; caveat — AdamC models were measured NOT reaching steady state even at 300 ViT epochs, so slow weight-norm drift at 100k steps is consistent with theory; (3) our cosine floor at 10% of peak sits on the paper’s recommended side (λ ∝ γ avoids terminal weight-norm suppression; non-zero terminal LR read as beneficial). ScionC’s headline gains are an optimizer-family swap — radar-only under startup velocity.

• 2026-08-09 lit-radar-0811 — two family-map entries: the head axis and the pretraining axis. (HiFlow page, 2603.27281): a third pole on the head-architecture axis — AR over scales (temporal-pooling ladder {1,2,4,8}), continuous flow matching per scale, no tokenizer. The citable datum: beats CARP (the VQ-token scale-AR twin, structure held fixed) 88 vs 85 avg and 90 vs 70 on threading — the cleanest controlled discrete-vs-continuous comparison on the tokenize-or-not question. Scale ablation saturates at 2–4 levels. No VLM trunk, ~104 NFE — family-map only. (VLA-JEPA page, 2602.10098): the predictive pole of representation supervision — trunk pretrained to predict a FROZEN V-JEPA2’s future latents (t+8, leakage-free: futures are targets never inputs), then a flow head fine-tunes. Payoff is robustness, not capability: LIBERO wash (97.2 vs 97.1 OFT) but LIBERO-Plus 79.5 vs 69.6, and the human-video share carries it (drop Something-Something: 79.5→62.9). Loses to π0.5 on task-level OOD — a trade, not an upgrade. Same integration point as Spatial Forcing (#11); wrong stage for our trained trunks; alive only if a trunk-pretraining arm ever opens.

2026-08-09 — lit 0812b: the head-axis map completes to four quadrants (DFM-VLA page, 2603.26320): discrete flow matching = discrete tokens + whole-sequence iterative refinement (16 steps, revisable tokens, metric-aligned tokenizer). CALVIN 4.58 / LIBERO 98.0 / LIBERO-Plus 77.8 vs π0.5 75.7; 2.4× AR decode with caching. With HiFlow the meta-lesson is now measured from both directions: commitment, not discreteness, is the expensive property — HiFlow holds structure fixed and swaps continuous-vs-quantized; DFM holds tokens fixed and swaps revisable-vs-committed; both wins point the same way, and our AR-trunk-vs-flow-expert panel gap is what that predicts. And the predictive-supervision pole gains its cheap self-anchored variant (OneWM-VLA page, 2605.07931): one pooled semantic token per frame, jointly denoised with actions under one flow objective, 14.7M LoRA on π0 — +10.4 LIBERO-Long, +40 pp real cloth-fold; monotone bandwidth sweep (1 token 53.1% → 12 tokens 20.5%) and the sharp scaffold ablation (unsupervised latent tokens 21.5% < no tokens 43.0% < supervised 58.1%). The pole now spans teacher-anchored (VLA-JEPA) to self-anchored (OneWM); the self-anchored end is the plausible entry for a trained trunk — aux rider, not architecture change. Regime caveat loud: LoRA-budget-scoped by their own admission.

2026-08-09 — lit 0813: the commitment axis gets its within-model intervention (AsyncVLA page, 2511.14148): HiFlow and DFM-VLA measured commitment between architectures; AsyncVLA measures it inside one — after a standard flow decode, re-noising just the low-confidence tokens and re-denoising them with trusted neighbors as context lifts SimplerEnv-Bridge 47.9 → 70.8, while doubling synchronous denoise compute buys only +3.2. Two thirds of the lift survives a coin-flip token selector (62.5), so revisability itself — not detection — carries the effect. Carried constraint: the mechanism must be trained in (partial-mask objective; bolting the two-pass inference onto a plain-flow model collapses to 7.3), and their σ_c=0.05 corrupted-context trick is the exposure-bias fix for any correction-conditioned module. Also (SA-VLA page): frozen VGGT-token injection = a fourth aux integration mode (read-only geometry, erosion-proof under RL), filed with the family on #11.

2026-08-09 — lit 0814: the adamc watch goes two-sided (Hyperball page, 2606.16899 + Anytime Pretraining page, 2602.03702): Hyperball (Stanford/Marin) derives the equilibrium law R⋆ ∝ √(η/λ) — AdamC’s λ ∝ η makes the norm target constant, so plateau-then-flat now has a third independent derivation (never citing AdamC); and its scale-invariance lemma (grad ∝ 1/‖W‖) adds the grad-norm side: if the correction holds, corrected-group grad norms should stay ~flat through decay too; a grad climb mirroring 1/√η with sagging norms is the uncorrected shape. Two free offline probes banked (per-matrix ‖∇L‖·‖W‖ constancy; stable rank), plus the sharpest interpretive trap yet: at λ=1e-5 on a pretrained 4B init the equilibrium may never be reached — flat norms could mean “decay inert,” not “correction working”; the grad side and the Muon-SW alignment cosine disambiguate. Anytime Pretraining adds a chart-note: cosine’s endpoint quality is largely the decay leg’s implicit averaging, so mid-run probe reads understate what the compute buys — the ladder ranks trajectories, it doesn’t price intermediate models. Also from the slice (X-Tokenizer): the commitment axis gains its zero-test-time-commitment corner — discrete tokens as pure training-signal (AR head disabled at inference, flow head executes, +8.25 long-horizon over FAST-as-auxiliary) — commitment, not discreteness, stays the expensive property from a fourth direction.

2026-08-09 — lit 0815: the adamc watch gains its failure-side frame and a measured disambiguator (Weight-norm criticality page, 2607.21005 + Weibull weight-scale page, 2606.19367): the fifth and sixth papers of the corrected-decay reading close the loop from both ends. Criticality (Xu group, SJTU): on scale-invariant layers decay shrinks norms unopposed while sharpness grows as 1/‖u‖² — halve the norm, quadruple the curvature — until norms cross a derived floor c⋆ = √(ηρ/2) and the loss spikes; in transformers the top Hessian eigenvector concentrates in the MLP blocks during spikes, and killing decay there alone both stabilizes and lowers loss. New named failure mode for the watch (“criticality approach”), a joint read on series we already record: per-group norm decline + that group’s grad climb (the 1/‖u‖ law again) + train spikes co-timed with the deepest dips. It also flips the decay-inert trap’s valence — every demonstrated spike lives at λ ∈ {0.5, 1} (187M) or 0.01–0.03 (toys), 4+ orders above our 1e-5, so flat norms at our λ are the safe corner. And a synthesis (ours, flagged): λ ∝ η is incidentally spike-protective — constant λ rides R⋆ ∝ √(η/λ) down toward the floor during LR decay; the correction pins the distance flat — a fourth independent reason the correction has the right sign. Weibull weight-scale (Ding, single-author): AdamW norm change decomposes into alignment/injection/decay forces — alignment is 88–94% of the budget during norm rise, and a cubic-spline displacement trick recovers it from sparse weights-only checkpoints at 92–94% accuracy (~2× the two-point baseline); decay needs no recovery (checkpoint norms × our analytically-known λ_t·η_t schedule, exact — the identity’s time-varying-λ requirement satisfied for free); injection is NOT recoverable weights-only (hook corrected) but is ~4% residual. The payoff probe for the ~20 banked 5k saves: per-matrix |F_decay|/|F_align| across the run — ratio ≪ 1 with sizable alignment = decay inert at λ=1e-5 (flat norms are alignment’s doing); ratio → O(1) into the cosine tail = the AdamC balance is real. Converts the trap from named ambiguity to measured number. Bonus: the spline-recovered ⟨W,û⟩ makes Muon-SW’s alignment-cosine probe computable from weights-only saves. Offline-probe list now: grad·norm constancy, stable rank, alignment cosine, distance-to-criticality margin (ρ_grad = ‖u‖²·gᵀHg/‖g‖², one grad + one HVP per group), force chronicle. Caveats: quote recovered forces, never forward-integrate norms across 5k gaps (15–24% error in their real-Pythia test); direct three-force ground truth validated only at 70M from random init — method transfer, not phenomenology, on a 4B pretrained trunk.

2026-08-09 — lit 0816: the adamc watch gains a plasticity frame (priced as analogy), and the WM-verifier pole a directional caution (WD-plasticity page, 2602.11137, Kakade group + FoMo-FD page, 2607.27511): the plasticity paper’s causal single-knob sweep: pretrain λ 0.5–1.0 beats the 0.1 default on post-finetune performance at 20 TPP (and base loss too — the trade-off only appears overtrained at 140 TPP, where λ=0.3 still wins downstream despite worse CE); pretrain loss is an unreliable proxy for finetune quality. Three correlational signatures: linear-probe separability up at every layer, W_QK pseudo-rank halved, train–val gap down. For the trunk axis: base benchmarks under-predict post-finetune quality — weight finetuned probes in trunk selection; the layer-wise linear-probe separability metric banked as a cheap pre-finetune plasticity instrument. For the adamc watch, a frame NOT a claim (hook corrected): λ∝η never appears in the paper, our λ=1e-5 is 4 orders below their range, and at λ≪0.1 they see no effect — the “AdamC preserves late-training plasticity for a stage-2 phase” story is a two-step analogy, record-only. FoMo-FD’s datum for the WM-as-verifier pole: on the same success-only action-conditioned latent WM, backward inverse-transport scoring detects 96.6% of failures where forward prediction-error scores 52.2% — the useful signal is post-hoc by construction (needs the observed endpoint), a caution against forward-sampling verifiers ranking candidate chunks pre-execution.

2026-08-09 — owner-requested MolmoAct2 deep dive (post, 2605.02881, AI2): the trunk axis gets its most actionable arm ever. AI2’s production VLA is built on OUR trunk family — Molmo2-4B finetuned into Molmo2-ER (specialize 20k steps on a 3.3M embodied corpus, then rehearse 1.5k steps re-mixing the original mid-training data, 50/50 best) — and the controlled ablation is the datum: swapping Molmo2 → Molmo2-ER under an identical discrete VLA lifts LIBERO-Long 77.6 → 83.6 (+6.0), larger than their per-layer-KV conditioning gain (+1.9) or K=8 flow samples (+1.75). Molmo2-ER is released (Apache-2.0, ~5.8k monthly downloads). The concrete arm: frozen-Molmo2-ER swap under the identical F recipe, panel delta vs F — externally priced, weights-in-hand, owner-decision. Also filed: their expert is 621M on a 4B trunk (15.5%; our F 9.2%, tiny 2.2%) — a production capacity anchor for tonight’s Δ_capacity read; and the per-layer KV result seconds our deep-tap direction.

2026-08-09 lit 0818 — a fourth attachment pole, and the plasticity watch loses a false companion (Qwen-RobotManip 2606.17846 + plasticity-at-scale 2606.24752): Qwen’s production recipe is a Qwen3.5-4B trunk joint-trained end-to-end (flow loss + λ=0.1 next-token VLM loss on a 9:1 robot:VL mix — the priced anti-forgetting recipe if we ever unfreeze) with a ~100M-order DiT expert on cross-attention to hidden states, alternating visual/language tokens per block — a topology distinct from our residual taps, MolmoAct2’s per-layer KV, and VLA-GSE’s spectral adapters; no frozen-trunk ablation, so no vote against our F result. Their key negative seconds VLM4VLA from the action side: standard benchmarks don’t separate pretrained from non-pretrained VLAs — only OOD suites do (LIBERO-Plus 89.0 vs π-0.5 84.4; RoboCasa365 35.9 vs 16.9) — one more argument for an OOD-style probe next to panel MAE. The plasticity-at-scale read is a hook correction for the adamc watch: its WD clause was a citation of 2602.11137 (already read), not new evidence — WD fixed at 0.1 throughout, zero grad-norm analysis, largest measured model 314M (4B claims are extrapolation); record-only. Its durable export is negative: dormant units, param norms, and attention entropy all FAILED to track plasticity onset even in-domain — never build watch/kill lines on those proxies; behavioral fixed-budget adaptation (our probe-MAE-under-budget shape) is the only instrument that worked.

2026-08-09 — lit 0819 (CL triangle, 2603.03818 + 2605.26820 + 2603.11653): the unfreeze recipes get their forgetting price list. The three-way “contradiction” dissolves in the tables — every paper shows zero-replay sequential FT forgets catastrophically on pretrained VLAs (the “surprisingly resistant” paper’s own zero-replay rows: NBT 0.56–0.76); what the VLM prior buys is a 2–4× better replay exchange rate and <10%-of-steps recovery, and the ablation pins it on the VLM prior, not action pretraining. Binding real-robot precedent for our regime: π0.5-2.7B full FT loses BWT −81.0 within 4k steps/task; episode replay ρ 0.02–0.2 at ~20% of batches fully fixes it (BWT +1.5) and beats joint retraining at matched compute. For the live vision-unfreeze question: vision-full-FT inside a constrained trunk is π0’s own default (the well-behaved config) — the vu5k arms stand. For any LANGUAGE/trunk unfreeze: pre-register LoRA on the LM + a replay-like anchor; our banked Qwen-RobotManip 9:1 + λ=0.1 LM-aux rider already exceeds the sufficient dose. Replay-free CL claims are scoped to LoRA + on-policy RL (both ablations collapse without either) — no license for replay-free offline BC.

Lit 0820 2026-08-09 (H2R emergence 2512.22414): second production datapoint on the pretraining axis — π0.5+ego’s 0% condition (base VLM only) shows human/cross-embodiment co-training pays ~nothing without diverse robot pretraining already in the trunk, i.e. VLM-benchmark inheritance ≠ robot-pretraining diversity (consistent with VLM4VLA’s proxy-collapse read). Together with MolmoAct2’s ER gain this says embodied trunk pretraining both improves action decoding AND unlocks data levers downstream — strengthening the live er_60k arm’s rationale beyond its panel delta: an ER-class trunk is the zero-cost way up their x-axis and the reopening condition for the #9 human-video lever.