JEPA/DSpark alternatives — what the 2026 record says about aux heads and draft heads

Date: 2026-10-16. Read-only: no config touched, nothing committed. A: are JEPA-style latent-prediction aux heads the right aux for a small byte-LM (9.2M–1B), or does a cheaper signal win? B: does a DeepSeek-style DSpark draft head (MTP + spec-decode) buy SPEED at train or generation time at ≤1B, given our byte level? Builds on, does not repeat: docs/papers/jepa.md (J1-J13), docs/decisions/2026-09-30-do-we-need-jepa.md (RETUNE verdict), docs/papers/dspark.md (D1-D16), docs/protocols/AB-PROTOCOL.md rows 1/1b, docs/decisions/2026-10-15-1b-readiness-plan.md §7. Method: ~14 search_papers queries (cs.LG/CL/AI/CV, 2025-01→2026-10) + ~30 get_abstract reads; every ID below resolved individually through the API. Marks: (A) from the paper’s own abstract · (M) abstract claim, not reproduced here · (ours) our measurement, source named · (S) SPECULATION, no source.

TL;DR

1. [A] No 2026 paper supports a latent-prediction aux on a byte-level or ≤100M AR LM, and the two on-regime studies point the other way: ProteinJEPA 2605.07554 (A/M) gains downstream “without improving MLM loss”, and from random init replicates at p=0.059; 2604.17228 (A/M) removes a JEPA-style gate and improves LM loss 3/3 seeds at 157.5M. Our own d8062d1 (ours) says the opposite (6.343 vs 6.425 BPB, 3/3). Both are true; they do not reconcile, and ours is the only direct AR byte-LM measurement either way.
2. [A] The signal to trust is NOT the aux loss. 2606.05753 (A/M): cosine/MSE alignment is negatively correlated with accuracy, r = -0.94, and the aux “reshapes the model via shared parameters rather than via the latent variable it nominally optimizes”; its prescription is a linear probe + corruption test — the instrument our protocol lacks.
3. [A] The cheapest competitor is already pre-registered at 2.1% of the model: the future-byte head (aux_fb_weight, AB-PROTOCOL row 1b, source 2404.19737 verified real), 197 120 params, no extra forward — so it does not pay JEPA’s measured price (the teacher halves the legal batch, benches/history.tsv:51-56, ours).
4. [B] The best ≤1B draft-head number in 2026 is 1.48×, on a different architecture family (IronLLM-0.6B, 2609.36860 A/M); the cheapest recipe is post-hoc (2610.00888 A/M: plain CE on ~2.5B tokens on a frozen target ≈ joint pretraining, 10³–10⁴× fewer MTP tokens). Neither is measured at our scale.
5. [B] Our exact topology has a published NEGATIVE. 2605.01106 (A/M): sequential hybrids (interleaved linear+attention = our KDA interleave) give draft acceptance α = 0.038 vs α = 0.68 parallel — 18×, scale-invariant (0.5B and 3B match).
6. [B] Our byte level is the one sourced advantage: 2610.05978 (A/M) — byte LMs learn segmentation-like structure, and exploiting it for speculative decoding yields 3.4× more accepted tokens than subword models.
7. Three brief items do not survive verification (§5): 2602.00232 is not “DeepSeek M2”; no LAUSE/LAUSE-N exists anywhere in the repo; “next-latency prediction” is not a term in use. 2607.05147 (DSpark) is real.
8. Nothing published shows train-time speedup from a draft head at ≤1B (§6) — the closest are RL rollouts (2606.12370, ≥1.5B-class) and AR image generation (2608.25386, 39% faster training).

1. A — the aux question

1.1 WHEN it helps

The 2026 record gives three necessary conditions, plus one diagnostic. Our arm satisfies none:

condition	source	do we have it?
a second view of the same content — its own authors call this “crucial”	2509.14252, restated as a substrate property by 2606.00091 (M)	No. A byte corpus has no text↔code / paraphrase / Q-A pairing. Unavailable, not merely unattractive.
an input bottleneck: student fed something the target is not	2212.07525 (M) — teacher on the unmasked sample, student on a masked one	No. Our mask selects loss positions; the predictor sees the full sequence (aux.rs:177, ours).
a target more abstract than the student’s own state	ProteinJEPA ablation: “adding shallower targets removes most of the task gain” (M)	No. Target = teacher’s T-average readout at the student’s own depth.
the target is load-bearing, not decorative	2606.05753 PRISM corruption test (M)	Never measured. No probe in the protocol.

2024-25 lore (EMA teacher + predictor ⇒ self-supervised win, imported from data2vec-2.0/DINOv2) is what the crate header inherited. 2026 evidence narrows it to: works with two views or a real information bottleneck; otherwise it degrades to lagged self-distillation. 2607.23531 (M) supplies the theory — MSE prediction of an arbitrary continuation is pulled toward the conditional mean of admissible continuations. 2608.05806 HiLP (M) adds that plain MTP (limited horizon) and plain NextLat (compounding rollout error) are each weak alone; its fix is a new mechanism (a hierarchical abstract latent), not a retune. The gap that matters: this pass again found no autoregressive natural-text pretraining A/B below 100M params, loss metric, 3 seeds, random init. That is why d8062d1 can neither be corroborated nor refuted from the literature.

1.2 WHAT signal tells you it worked

The literature’s answer changed in 2026 — the aux loss is explicitly not it: 2606.05753 (A/M) alignment r = -0.94 vs accuracy, latents “largely bypassed”, corrupting them shifts accuracy ≤4 points; ProteinJEPA wins downstream “without improving MLM loss”; 2604.17228 (A/M) measured at the loss and the aux lost 3/3. Our protocol already says the same from the other side (AB-PROTOCOL.md:397, ours): “Убывание aux-loss без выигрыша BPB или code-задач — не основание возвращать ненулевой вес.” Practical rule for dormouse: report held-out BPB (the protocol’s primary) plus a probe — a linear probe on out_acc for the target byte, and a corruption test (zero the aux path, re-score). If the corruption test says the contribution is not load-bearing, the term is ornamental whatever its weight or loss curve.

1.3 The CHEAPEST competitor, per published study

candidate	params added	per-step	per-token	VRAM	source
future-byte head (aux_fb_weight)	197 120 = 2.1% of small (ours)	one untied d_model→256 linear + CE, no extra forward	+1 shallow head	1 logits block	AB-PROTOCOL 1b; 2404.19737 (A)
LightMTP (latent MTP from own hidden states)	≤1% (M)	no conventional-MTP overhead, no external teacher	latent, no token head	not reported	2610.06031 (A/M)
FSP (future summaries)	an aux head (size unreported)	+1 head; reverse LM for the learned variant	summary target	not reported	2510.14751 (A/M)
JEPA + KoLeo (current default)	EMA teacher copy + predictor	EMA update 0.14–0.21% of a step (ours)	+1 full teacher forward	halves the legal batch: aux off = batch 16 @ 5908 MiB; teacher on ⇒ batch 4 OOMs (ours)	~/logs/ab8m_ab8m_iter4.log; benches/history.tsv:51-56
KoLeo alone	0	~0	~0	0	but over 10 rows (aux.rs:178, ours) — not the published estimator

The ranking is decisive because VRAM is our binding constraint (readiness plan §2.3). 2404.19737’s own abstract carries the caveat: “no overhead in training time” and “increasingly useful for larger model sizes” (induction-head evidence at 1–30M algorithmic tasks; LM gains at 7B/13B). The direct negative is 2505.22757 (A/M): “smaller language models (SLMs) struggle with the MTP objective” — its fix is a forward curriculum, not a bigger head. The byte-level anchor is 2608.16671 (A/M): on byte-level WikiText-2, 5 paired seeds, “Tested auxiliary feedback routes do not beat tuned backpropagation.”

2. B — the DSpark / MTP draft-head question

2.1 What is published at or near ≤1B

study	scale	speedup	assembly	VRAM
IronLLM-0.6B 2609.36860 (A/M)	654M	1.48× decode	X-MTP: shared-KV MTP + lightweight verification head, rollback-free; 6.2T-token pretrain then multi-domain on-policy distillation	not reported
2610.00888 (A/M)	frozen Qwen3-8B, K=3 heads	+12–16% (chain-aware verification), +11–14% (adaptive head-count controller)	post-hoc: plain CE on ~2.5B target-generated CoT tokens = 10³–10⁴× fewer than joint pretraining (MiMo-7B baseline)	not reported
2602.06019 (A/M)	not given	>3× decode at <5% GSM8K drop	self-distillation converts a pretrained NTP model in place; no auxiliary verifier	not reported
2602.01469 P-EAGLE (A/M)	GPT-OSS 120B/20B, Qwen3-Coder 30B	1.10–1.36×	parallel drafting; gradient accumulation within a sequence	not reported
2606.12370 Bebop (A/M)	Qwen3.5/3.6/3.7	up to 1.8× end-to-end async RL training	MTP + rejection sampling; end-to-end TV loss on the acceptance rate; pre-RL MTP training suffices	not reported
2607.24434 DraftExpert (A/M)	DeepSeek-V2-Lite, Moonlight-16B-A3B	1.45× decode	one draft expert/layer, self-distilled from the frozen MoE; expert-offload	drafter accelerator-resident by design

Three rows matter more than that column:

• 2502.09419 (A/M): on a frozen NTP model, hidden layers are “strongly specialized for NTP, making adaptation non-trivial”, and joint training “cannot fully overcome this barrier” — dormouse plans exactly that frozen-logits arrangement (aux.rs:259 frozen = logits.detach(), ours).
• 2605.01106 (A/M): the architecture-matched negative — isolating the linear-attention/SSM subgraph as a zero-cost draft gives α = 0.68 parallel / 0.038 sequential, 18×, scale-invariant; and it offers a free gate: a perplexity-degradation ratio predicts speculative viability without running SD (3.15× ⇒ α=0.37 at k=4; 81.96× ⇒ α=0.019).
• 2609.17184 LoopSpec (A/M): for looped transformers — our shape — “intermediate recurrent states can directly provide draft predictions without an auxiliary draft model”, up to 6.83×, training-free, optimal proposal depths in closed form. The strongest published analogue of our LoopBlock, and it needs no draft head at all.
• 2607.21535 Windowed-MTP (A/M): the draft’s full-context KV read “comes to dominate the draft cost” and “sharpens under hybrid/linear-attention targets”; windowing the draft’s attention only gives +28–44% per-decode-step at 1M, lossless. Relevant before any context extension.

2.2 Is “our byte level makes it exact replace-striking” true?

VERIFIED: the byte level removes tokenizer mismatch, and byte LMs are better speculation substrates — 2610.05978 (A/M): 3.4× more accepted tokens than subword. Also verified: in dormouse draft and target are one network (dspark.md D15: “the draft and target are the same network here”), so this is self-speculation, not a drafter/target pair. SPECULATION (S2): that this makes the speedup “exact” or lets the draft replace-strike the target at ≤1B in our shape. The published self-speculation result for our topology is α = 0.038; the published ≤1B number is 1.48× on a different model; not one study in §2.1 reports VRAM.

2.3 Our own blockers, before any DSpark A/B deserves GPU time

From docs/papers/dspark.md (ours, tree eeb3b73): D1 the window is one position shifted (model.rs:204 passes targets where Eq 4/5/6 need input_ids) — one-line fix, open; D9 the confidence head is wired without the Markov feature the paper makes load-bearing (aux.rs:311 passes None; with_markov has zero callers); D2 DSPARK_GAMMA = 4.0 is decoupled from dspark_k = 4 where the paper ties one γ=7 to both. Its own verdict: “Any A/B of the DSpark head right now would move three things at once and attribute the result to none of them.” Nothing in this pass changes that, and the port has no head at all (port/src/dormouse/config.py:163-165 is _unsupported(0.0, "DSpark draft head"), ours).

3. Table — candidate | cost | claimed gain | measured by | fit to our byte-MoE | plan-item ref

#	candidate	cost	claimed gain	measured by study?	fit to our byte-MoE	ref
1	pure CE (control; what the port runs today)	0	—	ours: BPB 3.001 @5k, window 81 920 B, 0 NaN	the baseline	AB-PROTOCOL row 1
2	future-byte head (aux_fb_weight=0.1, horizon 2)	197 120 params = 2.1%; no extra forward	induction-head dev at 1–30M; LM gains at 7–13B	2404.19737 (A/M): “no overhead in training time”, but “increasingly useful for larger model sizes”	best cost/gain available here; counter-evidence at 9M from 2505.22757	AB-PROTOCOL row 1b
3	JEPA + KoLeo (burn default, jepa_weight 0.05)	teacher copy + full extra forward; halves legal batch (ours)	ours: 6.343 vs 6.425 BPB, 3/3 seeds	ours d8062d1; literature: 2604.17228 negative 3/3, ProteinJEPA p=0.059 from random init (A/M)	works on our data, uncorroborated; pays the binding resource	AB-PROTOCOL 1; do-we-need-jepa
4	JEPA retuned (R1 norm + R2 input mask)	same as #3	none measured	no study; prescribed by 2212.07525 (M)	the honest version of #3 — but its metric is the one 2606.05753 flags	do-we-need-jepa §6.2 A2
5	KoLeo alone	~0	DINOv2 retrieval +8.3 mAP (A)	yes, but vision, n=thousands of patches	our estimator runs on 10 rows — not the same estimator	jepa.md J8
6	LightMTP (latent MTP, own hidden states)	≤1% params, no external teacher	parameter-efficient MTP	2610.06031 (A/M)	avoids a 256-way head; untested ≤1B	new §5
7	FSP (future summary head)	aux head + reverse LM (learned variant)	beats both NTP and MTP at 3B/8B	2510.14751 (A/M)	“predict something further” at the head, not the latent	new §5
8	HiLP (hierarchical abstract latent)	extra abstract latent + head	longer-horizon belief state; better SD efficiency	2608.05806 (A/M)	a new mechanism, competing with NextLat and MTP	new §5
9	DSpark draft head (as wired today)	RNN head + W₁/W₂; dspark_weight 0.1	γ=7 blocks; acceptance (paper’s 4B+ targets)	2607.05147 (A) at Qwen3-4B/8B/14B + Gemma4-12B — not ≤1B	3 stacked defects (D1/D2/D9, ours) + α=0.038 for our topology	dspark.md; AB-PROTOCOL 1
10	X-MTP (shared-KV MTP + verify head)	one head set	1.48× decode	2609.36860 (A/M) at 654M	only ≤1B draft-head datapoint; different attention family	new §5
11	post-hoc MTP head (CE on target CoT)	~2.5B extra tokens once, frozen backbone	joint-pretraining throughput at 10³–10⁴× fewer MTP tokens; +12–16%	2610.00888 (A/M) at frozen Qwen3-8B	cheapest recipe — but 8B and on a frozen NTP-specialized trunk	new §5
12	LoopSpec (draft from early recurrent states)	0 params, training-free	up to 6.83× inference	2609.17184 (A/M), looped transformers	closest published analogue of our LoopBlock; no draft head needed	new §5
13	component-aware self-spec (linear-attn subgraph as draft)	training-free	α=0.68 parallel / 0.038 sequential	2605.01106 (A/M), scale-invariant 0.5B→3B	NEGATIVE for our sequential KDA interleave	new §5
14	Windowed-MTP	training-free, drop-in	+28–44% per-decode-step @1M	2607.21535 (A/M)	long-context only; sharpens for hybrid/linear targets	new §5
15	EntMTP (entropy-guided draft depth)	training-free scheduler	throughput across entropy regimes	2606.27550 (A/M)	serving-side; we have no serving scheduler	dspark.md D13
16	MTP via self-distillation	one short pass on a pretrained ckpt	>3× decode at <5% GSM8K drop	2602.06019 (A/M)	the “no extra verifier” property is the interesting one	new §5
17	Bebop MTP for RL rollouts	MTP head + TV loss + rejection sampling	up to 1.8× end-to-end async RL	2606.12370 (A/M)	post-train lane only, not pretraining	new §5
18	EB-JEPA (energy-based JEPA)	256 candidate scores/position/iter	none in LM	2602.03604 (A/M) — a library (CV/video/world-models)	rejected: new objective, zero LM-pretraining evidence	do-we-need-jepa §3.2(d)
19	SOD (step-wise on-policy distillation)	teacher + per-step reweighting	fixes OPD’s cascade failure for TIR agents	2605.07725 (A/M) — post-training, not an aux loss	not a pretraining aux; belongs to the RLVR lane	agentic-post-training §TL;DR 3
20	circuit probe 2606.02378	0 (measurement)	maps induction/prev-token/BOS head emergence, 3 × 1B-class, 10 revisions each	2606.02378 (A/M)	an INSTRUMENT, not a signal — it cannot replace an aux; it is what our protocol lacks	do-we-need-jepa §2.1(3)
21	MTP curriculum (NTP→MTP ramp)	0 extra params	lets SLMs use MTP at all	2505.22757 (A/M)	the published precondition if any MTP head is tried at 9M	new §5
22	PRISM diagnostics (probe + corruption)	0 (measurement)	alignment vs accuracy r = -0.94; latents largely bypassed	2606.05753 (A/M)	the missing instrument for rows 3/4/5	new §5

4. Decision-digest

4.1 What goes into the port’s tests NEXT

Item 1 — the A/B: aux_fb_weight (future-byte head) vs the port’s current pure-CE recipe. Why this and not the JEPA retune: it is pre-registered (row 1b), cheapest by every costing we have (#2 vs #3: 2.1% params and no extra forward, against a term whose price is a halved batch), and the port already runs the control (pure CE, BPB 3.001 @5k, 1b-readiness-plan.md:479). Implementing the arm is the smaller diff.

• Budget: 3 seeds × 2 000 steps, the port’s landed WSD recipe. Matched: one batch . seq_len across both arms = the port’s current operating point, so the window (81 920 B) is identical; a different batch voids the table (AB-PROTOCOL §“Одинаковое число шагов ≠ matched-compute”).
• Fairness: equal processed bytes for the smoke, equal tokens for the verdict; print GPU-seconds per arm.
• Metrics in order: (1) held-out BPB at matched steps, window printed; (2) fb=<ran>/<asked> — 0/<n> means the horizon was past the sequence length and nothing trained (head did not run ⇒ void); (3) a probe — linear probe on out_acc for the byte at t+2 plus a corruption test (zero the head’s contribution, re-score) — row 22’s instrument, because the loss cannot answer the question.
• Decision rule (unchanged): win = arm mean BPB below the pure-CE mean by more than the spread across the control’s own 3 seeds; inside the spread = delete (ADR-0002).
• Pre-registered prediction, so it can be falsified: on 2505.22757 (A/M) and 2404.19737’s own “increasingly useful for larger model sizes” (A/M), the head lands inside the spread or worse at 9.2M. Write that down before running.
• Named asymmetry: the 100M ladder was dropped (owner, 2026-10-15 — b1 directly), so the only remaining rung is b1. A 2 000-step arm at 9.2M can falsify cheaply but cannot confirm a 7–13B effect. A loss or tie is evidence against carrying the arm into b1; a win is a licence to keep it, not proof it scales.

Item 2 — the test, not a race: the DSpark D1 window gold vector. dspark.md §7 test 1 (x[i] = i; assert the RNN head at (p, s) is fed x[p+s] and supervised on x[p+s+1]) fails on the current tree, needs no GPU, and gates any future draft-head decision; D9’s width assertion (d_model + markov_rank) rides the same file. Cost: CPU only. Do not spend a seed set on the DSpark A/B while D1/D2/D9 are live (§2.3).

4.2 What rides “measured later”

deferred measure	why deferred	trigger
JEPA retune arm (R1+R2)	#3 already won on our data; the retune’s prescribed metric (cosine) is the one 2606.05753 flags at r = -0.94; its cost is the binding resource	only if the future-byte head loses and the aux family is still wanted — then run it with the probe, not the loss
KoLeo re-pooling (per-position, not mean_dim(1))	not the same estimator as the published one at 10 rows	only alongside the JEPA arm; never standalone
correct@budget(bytes) for the draft head	the post-train lane does not exist yet; it is the metric that would show “replace-striking” is real	first RLVR report after SFT (post-training.md; readiness plan §7.2)
byte-level acceptance rate α on our decode path	no serving path yet; 2605.01106 gives a free predictor worth measuring first	when port/src/dormouse/model/decode.py grows a verified-draft path
LoopSpec-style drafting from recurrent states	needs the loop’s intermediates exposed as a draft; training-free, cheap later	after b1, if generation speed becomes binding (not training)
Windowed-MTP	long-context only	only if context exceeds the current window

5. New IDs verified (5 + 2, and the corrections)

id	date	authors	one line
2606.05753	2026-06-04	Zhang, Fang, Liang	“Cosine Misleads” — cosine/MSE aux alignment inverted vs accuracy (r=-0.94); PRISM probe + corruption test; the aux reshapes the model via shared params, not the latent.
2610.00888	2026-10-01	Badarayani, Jay, Zhou, Julienne	Post-hoc MTP heads: plain CE on ~2.5B tokens on a frozen target matches joint pretraining at 10³–10⁴× fewer MTP tokens.
2605.01106	2026-05-01	Borobia, Seguí-Mas, Tormo-Carbó	Component-aware self-speculation: linear-attention subgraph as a zero-cost draft ⇒ α=0.68 parallel / 0.038 sequential, scale-invariant. Our topology is the loser.
2609.17184	2026-09-15	Cho, Cui, Kim	LoopSpec: looped transformers draft from early recurrent states — training-free, up to 6.83×, closed-form optimal proposal depth.
2610.05978	2026-10-05	(cs.CL)	Byte LMs learn segmentation-like abstractions; exploiting them gives 3.4× more accepted tokens than subword in speculative decoding.
+2610.06031	2026-10-05	Czinczoll, Kallini, de Melo et al.	LightMTP: latent MTP bootstrapped from the model’s own hidden states, ≤1% added params, no external teacher.
+2505.22757	2025-05-28	Aynetdinov & Akbik	The direct negative for #2: SLMs struggle with the MTP objective; a forward NTP→MTP curriculum is the published fix.

Corrections to the brief (verified, not assumed): 2602.00232 is NOT “the DeepSeek M2 paper” — it resolves to “Complexity of Quantum Trajectories” (quant-ph, 2026-01-30), is absent from docs/papers/, and has no provenance.tsv row. LAUSE / LAUSE-N: nothing. A repo-wide grep (md/tsv/toml + .bulba/) returns nothing and arXiv has no match — the ID is not real, as suspected. “next-latency prediction” returns zero results; the real family is next-latent / NextLat (2608.05806, 2607.00958, and the negative 2609.36227, “One-Step Next-Latent Prediction Is Not a World Model”). 2606.02378, 2605.07725, 2602.03604, 2604.17228, 2404.19737 all resolve and are as described. MTP sets, by author (the 2511–2610 wave and its roots): (pretraining-aux) 2404.19737 Gloeckle, Youbi Idrissi, Rozière et al. · 2505.22757 Aynetdinov & Akbik · 2510.14751 Mahajan, Goyal, Youbi Idrissi et al. · 2610.06031 Czinczoll, Kallini, de Melo · 2610.05978. (draft head / self-spec) 2502.09419 Mehra, Alonso Garcia, Mauch · 2602.06019 Kirchenbauer, Hans, Bartoldson et al. · 2602.01469 Hui et al. · 2606.27550 Chen · 2607.21535 Valliappan · 2610.00888 Badarayani et al. (RL rollouts) 2606.12370 Li, Jiang, Xu · 2605.28184. (looped / hybrid self-spec) 2609.17184 Cho, Cui, Kim · 2605.01106 Borobia et al. · 2607.24434 Han. (DeepSeek line) 2607.05147 Cheng et al., DeepSeek-AI + PKU (DSpark) · 2606.26744 HyperDFlash (DeepSeek-V4 MTP).

6. VERIFIED (sourced) vs SPECULATION (my inference)

VERIFIED. Every ID in §3/§5 resolved through the API with the date shown; numbers are the authors’ abstract-level claims unless marked (ours). Ours: d8062d1 6.343 vs 6.425 BPB, 3/3 seeds (docs/guides/ab-testing.md:138); JEPA EMA 0.14–0.21% and the halved legal batch (benches/history.tsv:51-56); future-byte head 197 120 params = 2.1% of small (AB-PROTOCOL 1b); port k5000 BPB 3.001 @ 81 920 B, CE 1.46 (1b-readiness-plan.md:479); port JEPA/DSpark/aux_fb all _unsupported (port/src/dormouse/config.py:163-165); DSpark D1/D2/D9 open (docs/papers/dspark.md); 2602.00232 is quant-ph. SPECULATION (mine, no source). S1 that the future-byte head is the right first aux A/B — an inference from cost and pre-registration, not from a scaled result; the one on-regime citation predicts it underperforms at 9M. S2 that our α would land near 2605.01106’s 0.038 — our loop is weight-shared with a per-iteration router, closer to a sequential hybrid than to a looped transformer’s per-depth states, but that is a topology argument, not a measurement. S3 that byte-vocabulary identity makes the draft/target pair cheaper here; supported in spirit by 2610.05978 (A/M), never measured on our decode path. S4 that a probe + corruption test would show our JEPA term is not load-bearing (the reading do-we-need-jepa §3.1 argues for) — the instrument is published, the result on our model is not. I cannot answer on available information: (i) whether any 2026 study shows train-time speedup from a draft head at ≤1B — the closest are 2606.12370 (RL rollouts, ≥1.5B-class) and 2608.25386 (AR image generation, 39% faster training, A/M); (ii) per-study VRAM cost for any draft head — not one study in §2.1 reports it; (iii) independent support for d8062d1 at our regime — none exists, and ours remains the only direct evidence either way.

7. Open questions

1. Does a corruption test show our JEPA+KoLeo contribution is load-bearing on out_acc? Cheapest possible experiment: one checkpoint, zero the aux path, re-score BPB.
2. What is our perplexity-degradation ratio (the 2605.01106 free predictor)? If it is bad, the DSpark lane closes without a GPU run.
3. Does LoopSpec’s draft-from-early-states beat a trained draft head here? Training-free, so nearly free to test once the loop exposes intermediates.
4. Is 2610.00888’s post-hoc recipe the right shape for a byte drafter, where “CoT” is not a natural unit? No source. (S)
5. docs-site/tools/manifest.mjs needs a line for this document before npm run check is green — not done here (one-file mandate, no commit).

8. Sources

arXiv, fetched 2026-10-16 via the API (id · date as returned): 2404.19737 2024-04-30 Gloeckle et al. · 2502.09419 2025-02-13 Mehra et al. · 2505.10518 2025-05-15 · 2505.22757 2025-05-28 Aynetdinov & Akbik · 2510.14751 2025-10-16 Mahajan et al. · 2602.01469 2026-02-01 Hui et al. · 2602.03604 2026-02-03 · 2602.06019 2026-02-05 Kirchenbauer et al. · 2603.19714 2026-03-20 Tang et al. · 2604.17228 2026-04-19 · 2605.01106 2026-05-01 Borobia et al. · 2605.07725 2026-05-08 · 2605.07554 v2 ProteinJEPA (via do-we-need-jepa) · 2606.02378 2026-06-01 · 2606.05753 2026-06-04 Zhang, Fang, Liang · 2606.12370 2026-06-10 Li, Jiang, Xu · 2606.27550 2026-06-25 Chen · 2607.21535 2026-07-23 Valliappan · 2607.23531 2026-07-26 · 2607.24434 2026-07-27 Han · 2608.05806 2026-08-06 Shi, Pearce, Tomar, Sen · 2608.16671 2026-08-17 Murugan · 2608.25386 2026-08-26 Niu et al. · 2609.17184 2026-09-15 Cho, Cui, Kim · 2609.36860 2026-09-29 Yang et al. · 2610.00888 2026-10-01 Badarayani et al. · 2610.05978 2026-10-05 · 2610.06031 2026-10-05 Czinczoll et al. · 2602.00232 2026-01-30 not DeepSeek M2 (quant-ph). Repo: docs/papers/jepa.md, docs/papers/dspark.md, docs/decisions/2026-09-30-do-we-need-jepa.md, docs/protocols/AB-PROTOCOL.md:216-217,384-397, docs/decisions/2026-10-15-1b-readiness-plan.md:477-479, docs/guides/ab-testing.md:138, benches/history.tsv:51-56, port/src/dormouse/config.py:163-165, port/src/dormouse/model/model.py:112-122.