Pre-registration DRAFT: F-then-joint — is the frozen expert’s capital a key to joint training? (#4)

2026-08-09 ~14:3xZ. DRAFT, not yet immutable. It becomes the binding pre-registration when three things happen, in order: (1) the instrument section lands oracle-gated (check.py green), (2) the owner OKs a box window for it (it competes with the adamc_100k endpoint’s own stage-2 work, see Scheduling), (3) the execution queue item is created owner-visible. Edits before finalization are edits; after, amendments per the standing rule. Ideas #4; basis: the stage-2 attachment decision memo (frozen default adopted; KI-joint closed-unmeasured at a measured 4.11× step cost), which explicitly unblocked this draft and set its burden: any joint-flavored escalation must argue against that measured cost. The argument is below, before the recipe.

Question

The attachment screen asked whether a trunk adapting from the start of expert training (KI-joint, random-init expert) beats a frozen trunk. It never asked the question APT (2606.12366) says is the right one: once a flow expert has been trained to convergence against the frozen trunk, does then unfreezing the trunk — flow gradients flowing in, no stop-grad — buy anything a frozen continuation doesn’t?

APT’s diagnosis: the seam damage that KI-style guards defend against is caused by the expert’s random initialization — an uninformed head learns the vision-shortcut and its noisy early gradients wreck the trunk; with a pretrained expert, their best row unfreezes everything with no gradient stopping at all (+8..+26 pts over the frozen row in their grid; “stop-gradient is not a necessary condition”). LP-FT supplies the mechanism theorem (feature distortion is front-loaded while the head is uninformed — align the head first, then unfreeze) and ActionX the same shape independently (+38 pts LIBERO-Long over joint-from-scratch, Frontiers caveat carried). Two production stacks (RDT2, Qwen-VLA) also train the expert against a frozen/protected trunk first — though RDT2 then never unfreezes at all, which is the other way this rung can land.

We hold the exact substrate this literature names: the screen’s F arm is a converged flow expert on a hard-frozen action-pretrained trunk — APT’s Stage 1, banked and paid for. This rung spends that capital.

Why this is worth proposing against a measured 4.11× step cost

The memo’s burden, answered head-on:

1. The 4× objection prices a lineage, not a rung. The F-then-joint shape is precisely “cheap frozen phase long, joint phase short”: if it works, the joint phase is a bounded final segment (here 5k steps against a 10k+60k frozen history), not a 4×-forever recipe. LP-FT’s compute-Pareto case is this shape.
2. The rung is the cheapest remaining discriminating contrast. K-from-noise answered “does trunk adaptation help an uninformed expert” (no, in 4k matched steps). It did NOT answer the initialization question — APT’s grid says those are different regimes with opposite verdicts. One bounded rung closes the whole joint direction measured either way; without it, “closed unmeasured” stays a caveat on every future stage-2 pre-reg.
3. Committed cost is capped at ~32 GPU-h (5k matched arms + evals, ceiling 35), under half the screen’s 70 ceiling. The 10k extension spends only on a positive CI (see decision rule).
4. The payoff side is not small. If the joint phase buys even a 0.3-class panel gain, that re-ranks the stage-2 recipe for every future trunk in this class — including the adamc_100k endpoint’s full-length attachment, which per the memo currently binds to frozen.

Arms — the joint bundle is the contrast, initialization is held

Both arms warm-start from the SAME parent: the screen’s F endpoint fontaine_molmo2_flow_frozen_10k_ddp4/step_010000 (trunk = the 60k AR endpoint verbatim — F never moved it; expert = the converged h1024×12 flow expert, panel 9.4157, probe 9.38@10k still trending down). Matched +5k steps, eff-48, identical data order.

• F2 exists because F was still improving. J@+5k vs the banked F@10k number would credit the joint phase with plain extra training. The paired read is J@+5k vs F2@+5k: same parent, same +5k data, the joint bundle the only difference.
• J is the APT best-row analog, not another K. Flow gradients DO enter the trunk through the taps (the thing the screen’s guard refuses for random-init runs — see Instrument). The CE rider stays on, phase-1 verbatim (α=1, same aux fields/dropouts, tables continuing from the 60k endpoint’s expert.safetensors), for the Wall-OSS reason (co-train > flow-only in their from-scratch grid) and because it keeps the CE-health watch and drift read interpretable. J’s step is K’s step minus stop-grad plus tap gradients — so K’s measured 3.782 s/step anchors the cost projection (~4.0 s/step assumed; the rate gate measures).
• Changing {trunk trainable, flow-grads-in, CE rider} together is deliberate: the rung tests the joint phase as a recipe. If it wins, attribution gets its own pre-reg; if it loses, the bundle’s best published form lost, which closes the direction.

Shared recipe constants (identical across arms, NOT under test)

• Start: --init-from the F@10k checkpoint (weights only, fresh optimizer, step 0) — J via the materialized composite (Instrument §1). Warmup 500 both arms (fresh optimizer state).
• Fresh shuffle seed, standing owner rule: --seed 2 (new vs phase-1’s 0, F’s 0, adamc’s 1) — the SAME new seed both arms; matched data order is what makes the probes and the panel read pair. --holdout-episodes 0.1 --split-seed 0 unchanged (the split is identity, not shuffle).
• Surface: residual taps 12 @ stride 3 (layers 2,5,…,35), expert h1024/12, adarms, bidirectional, --decoder-lr 1e-4, --max-crops 1 — the screen’s pinned surface, verbatim.
• Topology: box 4×H100, eff-48 (12/rank), --zero1 --backward-chunks at the K-smoke values (B12c6), --activation-checkpointing for J (K’s prerequisite carries; F2 runs without, as F did — memory, not semantics). Eval 256 @ every 500, save every 1250 (async), both arms.
• J’s unfreeze surface: phase-1 verbatim — backbone-text-lr 2e-5 --grad-clip 100, frozen embeddings/lm_head (the molmo2 unfreeze surface), --prompt-generate-bracket, same --aux-fields and dropouts.
• Run names fontaine_molmo2_flow_fcont_5k_ddp4 / fontaine_molmo2_flow_fjoint_5k_ddp4; sequential, F2 first (cheap arm shakes out the warm-start path); babysit.toml entries at each launch, first-poll util+rate check per standing rule.

Instrument (to land, oracle-gated, before finalization)

Status 2026-08-09 ~15:0xZ: §1–§3 LANDED, oracle-gated (check.py 596 green, +12 new oracles in tests/test_fjoint_init.py). The spellings: §1 = fontaine/scripts/materialize_fjoint_init.py (trunk-coherence byte-guard included — a wrong phase-1 checkpoint is refused, not composed); §2 = --joint-unfrozen-seam (warm-start-only: requires --init-from, contradicts --seam-stop-grad, refusal verbatim-preserved without it; the launch banner prints seam UNFROZEN (flow grads enter the trunk)); §3 verified against the real writer on the fixture family. §4 (memory smoke) is box work at launch, per its own text. Finalization now waits only on conditions (2) owner go + (3) the queue item — both at the sequencing decision.

Semantics frozen here; flag spellings are implementation’s:

1. Composite warm-start materializer (audit result: no train.py surgery needed). --init-from F@10k --joint-ce correctly aborts today — F’s checkpoint carries no joint_ce.safetensors, and --backbone-init-from would build the expert fresh, discarding exactly the capital this rung spends. A small script (the materialize_joint_ar_view.py precedent, inverse direction) copies F@10k and adds joint_ce.safetensors := the 60k endpoint’s expert.safetensors (phase-1 FAST tables, continuing not restarting) + the config section. Oracles: expert bytes ≡ F@10k’s; rider bytes ≡ the phase-1 tables; --init-from --joint-ce load round-trips strictly.
2. Naive-joint guard escape, narrowly scoped. The train.py guard refusing --joint-ce without --seam-stop-grad (“a published collapse (KI), refused as a run”) is CORRECT for random-init and stays. A new opt-in flag admits the combination; the guard’s refusal must still fire without the flag. Oracle: under the flag, flow-loss gradients into trunk parameters are nonzero (the existing negative-control oracle becomes this run-mode’s positive contract); without it, the parser error is verbatim-preserved.
3. Drift-read compatibility: materialize_joint_ar_view.py accepts J’s checkpoints (same file shape as K’s) — verified against the real writer on the fixture family before launch.
4. Memory smoke, J config exact: one 150-step B12c6 rung (smoke_attach_k_ddp4.sh pattern), pass = rc 0 AND vram_alloc_peak_gib ≤ 71. K’s 57.34 green does NOT automatically bind — J adds tap-gradient backward paths K detached. Red ⇒ matched downshift both arms, loudly echoed.

Gates (in-run, mechanized where precedent exists)

• vram_alloc_peak ≤ 71 GiB, both arms (standing box rule).
• Rate gate, measured not judged (attach_rate_gate.py pattern, first jsonl window): projected 5k-phase batch total > 35 GPU-h ⇒ kill, rung closes incomplete, owner steer (no downshift branch — 5k IS the short form; halving it guts the read).
• Kill bars, J (K1-style): NaN/inf ⇒ kill. Probe > 12.38 (= F endpoint 9.38 + 3.0) at any eval ⇒ kill at next save boundary — the trunk-damage backstop if APT’s regime claim is wrong here. Probe − F2’s matched-step probe > +1.0 at any eval ≥ 2000 ⇒ kill (the joint phase is actively hurting; F2-first scheduling banks F2’s curve before J needs it).
• CE-health watch (record, not gate): J’s loss_aux vs the screen’s banked phase-1 tail anchor (~3.68) at every eval.

Frozen reads

Panel: plans/holdout_curated_v0_k4l2_panel_v2.json, flow keying heun30/draws1/stable, 4-GPU sharded, sha256-pinned — the screen’s spec verbatim. Paired per-frame from --dump-predictions npz, seeded bootstrap 95% CI (seed 0, 10k resamples). attach_seam_results.py machinery reused with explicit stems (J in the K slot, F2 in the F slot — the paired-read code is arm-name-agnostic; verified on its existing oracles before the read).

1. Primary: Δ_joint = chunk_mae(J@+5k) − chunk_mae(F2@+5k), paired CI. This one number is the rung.
2. Decision rule, frozen now:
   • Δ_joint ≥ 0 or CI includes zero ⇒ the joint direction closes measured for this trunk class — from-noise (K) and from-capital (J) both null; the frozen default’s standing becomes a measured fact, not a priced-out default. No extension, no refits, no α fishing.
   • Δ_joint < 0, CI excludes zero, drift band respected ⇒ the 10k extension fires: both arms continue to +10k (same rates, ~+31 GPU-h, global ceiling 70), panel read repeats. Adoption bar at +10k: Δ_joint ≤ −0.3 with CI excluding zero AND drift band held ⇒ the full-length stage-2 recipe for this trunk class gains a bounded joint final phase; the adamc_100k endpoint attachment pre-reg cites this rung (its frozen phase is unchanged — this appends, the memo’s binding is amended not overturned). Won at 5k but under the bar at 10k ⇒ recorded, frozen default keeps, direction closed measured-small.
   • J wins but breaks the drift band ⇒ wins-with-named-cost: AEGIS orthogonal-projection repair un-banks as the named escalation; adoption waits for owner steer.
3. Execution oracle: both arms ≥ 1.0 chunk-MAE below state-copy 11.7639, else the rung is VOID, not negative (screen read-3 rule verbatim). Context anchors quoted-never-deciding: F@10k 9.4157; K’s probe curve; gemma lineage 6.5997 (cross-trunk, directional).
4. Trunk-drift (J only): greedy AR panel of J’s materialized AR-view @+5k (and @+10k if extended) vs the 60k endpoint 5.8602, band |Δ_AR| ≤ 0.3 inclusive — the screen’s read 4 inherited with its comparator.
5. first_mae mirrors of 1 and 3; per-step-in-horizon curves both arms (record-only).

Numbered expectations (banked before data)

1. Both arms beat state-copy decisively and neither regresses above F@10k’s 9.4157 — confidence high; an F2 regression voids the rung (warm-start or seed pathology, not a seam fact).
2. F2@+5k improves on 9.4157 by roughly 0.1–0.25 — the 60k continuation’s slope analog, F’s probe was still falling — confidence medium.
3. Δ_joint is the genuinely open number. APT/ActionX say negative and large; K’s matched-probe null plus the Wall-OSS reading (phase-1 CE already routed the action gradients) say ~0. Banked: CI-excluding-negative at +5k gets confidence medium-low — this rung exists because the literature and our own trunk’s evidence point opposite ways.
4. J’s drift stays in band — CE co-training anchors the trunk, and LP-FT’s mechanism says the distortion channel is disarmed once the head is informed — confidence medium.
5. Falsified if Δ_joint ≥ 0 / CI spans zero: the initialization escalation closes for this class, and every future stage-2 pre-reg cites a measured joint null from both starting points.

Cost & scheduling

Committed (5k phase): smoke ~0.2 + F2 ~5.1 (0.92 s/step measured) + J ~22.2 (4.0 s/step assumed from K’s measured 3.782) + 2 panel evals ~2.5 + drift AR panel ~1.7 ≈ ~32 GPU-h, ceiling 35. Conditional extension: ~+31 ⇒ global ceiling 70 (spent only on a positive 5k CI). Weights: F@10k is on the box + its expert capital in fontaine-checkpoints (backbone dedup’d to the 60k trunk); execution re-verifies both before the materializer runs.

Venue: box 4×H100, opens no earlier than the adamc_100k endpoint + its chained panel (~08-12 ~17:00Z+). Sequencing question for the owner at finalization: this rung informs the adamc endpoint’s own stage-2 attachment (information-optimal order: rung first, then the full-length attach cites its verdict), but the attach is the deliverable — owner go decides which takes the box window. The scale caveat (2606.14153) is carried in both directions: the verdict is a Molmo2-4B-class fact; a different trunk class re-screens.