geolip-aleph-qwen-3.5-0.8b-instruct

This is the research record. If you're here for something usable:

you want	go to
a working caption adapter for Qwen3.5-0.8B (validated both seeds, standalone loader)	qwen3.5-0.8b-relay-caption
the framework — train / attach / align / detach these adapters yourself	amoe-lora
the full experimental story with citations	`article_ft3.md` in this repo
the map of every related repo	`REPO_MAP.md` in this repo

The scientific headline (n=48 matrices; 2-seed on the replicated verdicts): always-on single-task adapters on this trunk are mutually destructive — the depth specialist alone reduces captioning to token-F1 0.0014 with zero terminations — and the aleph dispatch rescues specialize-regime anchors almost completely (caption returns to within 0.02 of gauge under every specialist mask) while failing blend-regime anchors entirely (the math experts escape damping, ratios 0.86–1.6 vs the ≥3× target, and measurably trample chain-of-thought the trunk handles natively). Regime — not machinery — decides whether dispatch helps, and the probe layer reads the regime in advance (directionally) for the image family. That one law is measured at four scales across exp011/014/017/018.

Reading order: the zero-shot fingerprint below → the experiments table (each row links a self-contained package with raw ledgers and self-asserting results) → the dated amendment blocks inside exp005–014 and exp021. The amendments, a formal retraction, and a corrected overclaim are not blemishes; they are the method — every package was adversarially audited (several retroactively), and the audits' corrections shipped as dated blocks inside the packages.

Second-model validation of the geolip-aleph-qwen program on Qwen/Qwen3.5-0.8B — a dual-tower VLM whose LLM tower is a hybrid: 18 Gated-DeltaNet linear-attention blocks + 6 full-attention blocks (indices 3,7,11,15,19,23), d=1024, 248,320-token vocab, tied embeddings. Every law certified on the Qwen2.5-0.5B line is re-tested here under one discipline: preregistered rules where applicable, strong zero-shot baselines (this is a post-trained model), single-seed results labeled as candidates, honest negatives shipped as findings. The campaign's full runtime ledger is campaign_ledger.jsonl (every run row, minute 0 to close).

The dispatch under test is the aleph closed-form addressing M̂ = Σₖ sinh(uₖ)Aₖ / Σₖ cosh(uₖ) — dense signed mixing, frozen key path, no selectors, no top-k, no load-balancing losses.

Substrate (frozen at the gate)

fp32 throughout (TF32 off), gradient checkpointing + chunked CE (248K-vocab logits never materialized per-sequence), memory fraction 0.92 on a 48GB card. Gates passed before any science: template constants discovered (not assumed) — <|im_end|> / <tool_call> identical to the 2.5 line; control vocab bands (30000–33063) collision-free against the 248K tokenizer; adapter wrap fires on both DeltaNet and full-attention blocks; fp32 DeltaNet fallback path benched at 3548 tok/s prefill; a 9k-token training step fits in 16.22GB under the riders. See substrate/v35_substrate.py.

Experiments

#	package	status	headline
001	`exp001_placement/`	complete (s0)	all24 placement = series default; no runaway SSM-drift within 9k (norm ratio +2.3→+4.0%, still rising at window edge; s0); gate-growth did not replicate (gates shrink, ppl −41%)
002	`exp002_registers/`	complete (s0+s1)	register capture transfers; seed lottery replicates; zero-shot is strong (0.73–0.92); always-on stack costs +9.2/+9.9 ppl@512 (s0/s1)
003	`exp003_termination/`	complete (s0, n=8) — candidate	termination direction favors rows (7/8 vs 6/8 — a one-sample gap; the direction is carried by the failure mode: a capped packed sample degenerates into repetition); the instruct prior softens the sharp dichotomy; includes the judge-window amendment
004	`exp004_caption/`	complete — 2 seeds	caption anchor: token-F1 0.408 → 0.706 (s0) / 0.401 → 0.704 (s1) — the +0.30 distribution-match delta replicates within 0.005 across seeds (no seed lottery); terminate/nonempty 1.0; the deliverable's base checkpoint
005	`exp005_specialists/`	complete (s0) — candidates	bbox: 0.000 → 0.69 valid / 0.89 IoU-score (genuine grounding gain); depth: valid JSON scoring 0.0 → 0.83 pair-order accuracy (cleanest gain); fixation: a format anchor (0.35→1.0 valid; localization was already 0.94 when format succeeded); first-pass judgments had been invalidated (GT contract) and are preserved with reasons
006	`exp006_math/`	complete (s0) — negative	substrate already at ceiling on synthetic math (0.958–1.0 zero-shot); training gains nothing and costs ppl 24.9 → 116.5; the causal amendment measured the mechanism: indiscriminate runtime amplitude (~90% of on-task delta ratio on neutral text) compounding superlinearly with depth; detach counterfactual recovers the bare trunk exactly (24.3914); math excluded from the collective roster as trained
007	`exp007_collective/`	complete (s0) — candidate	five anchors under damped dispatch cost +11.0 ppl — roughly the price of the exp002 register stack, but ~3× the exp009 multi-task monolith (+3.66, same four tasks) (see the tax ladder below, and its limits); 5,000 steps with zero starvation alarms; blend-dominant usage (caption 0.77); the trainable `new` stays alive at 0.03 with no dedicated stream (exp010 later showed that survival is not benign: composites collapse with `new` active and recover on its removal; candidate, s0)
008	`exp008_battery/`	stage 1 complete (s0, n=12/cell) — candidates	surgical decoupling replicates multimodally: three double dissociations (caption-off, bbox-off, depth-off each kill exactly their own task; depth's is a pure content kill at intact validity); fixation-off changes nothing — the redundancy the P1 blend call predicted; `new`-alone reverts everything to ~frozen (no absorber — and exp010 subsequently closed the controller reading against `new`: all_on 0.0 on composites, spec_only rescues to 0.417/0.167; candidate, s0)
009	`exp009_monolith/`	complete (s0) — candidate	single always-on multi-task stack (no dispatch) matches the collective on bbox/fixation/depth, loses caption 0.588 vs 0.743, taxes +3.66 ppl vs the collective's +11.0 — the cheapest measured carrier; composites 0.0/0.167
010	`exp010_controller/`	complete (s0, n=12/cell) — candidate	controller hypothesis not supported: all_on − spec_only = −0.417/−0.167 vs the preregistered ≥+0.15; mechanism classified by the raw-dump amendment: format derailment — with `new` active, 9/12 detect samples ramble past budget without closing JSON and 12/12 fixate samples emit no JSON at all, while spec_only makes clean two-part attempts under the identical judge; surviving positive: emergent composition from dispatched specialists (spec_only 0.417) beating the monolith (0.0/0.167)
011	`exp011_probes/`	complete (s0) — candidates	toggle law confirmed bit-exact (all-off ≡ frozen, max
012	`exp012_weakness_scan/`	complete (s0) + 2 amendments	the weakness map (10 math domains × 3 tiers × 2 formats + 3 code, zero-shot): control reproduces the exp006 ceiling at 1.0; nine domains in the 0.10–0.60 learnable band; format dissociation splits weakness into format-weak (CoT rescues direct floors: algebra t2 0.0→0.96) vs genuinely weak (both low: arith t3, fractions t2); code = calibrated negative (fix solved at 1.0; write a genuine floor on the repaired harness; predict low, ≤0.25 across two re-judges); amendments: predict judge-bug rejudge (e012b), CoT-window recheck with sigil rates (e012c — truncation was large: numtheory t3 0.67→1.0)
013	`exp013_experts/`	complete (2 seeds on the clean wins + held-out re-judge + wall retry)	independent AMOE-LoRA experts on scan-selected weaknesses. Held-out verdicts: algebra-steps generalizes at ceiling (1.00/1.00) on provably-unseen questions (format-pair win: steps ≫ direct 0.79/0.57 — direct-format elicitation of a CoT-latent capability); fractions clean (+0.33); sequences/baseconv gains are memorization — their band-tier question spaces (480/248) are smaller than the training draws (space-exhausted: no held-out set exists). Seed robustness: both clean wins replicate at s1 (algebra-steps +0.7917 vs s0 +0.7916; fractions +0.375). Capability-wall retry: the direct arith expert's +0.04 noise becomes 0.208 → 0.417 with derived-steps targets (t2 regresses at both seeds, −0.21/−0.25; s1 replicates the break: 0.333→0.458) — and the campaign's first positive off-domain spillover: algebra t1 rises +0.375 under the always-on arith-steps anchor (the stepwise emission format transfers across domains)
014	`exp014_basins/`	complete (s0 + direct-algebra control + e014c composite retry)	5 frozen experts under one aleph dispatch, keys-only training (7,680 params), preregistered P5a–d: P5b is seed-fragile (s0 keys: holds on all band cells vs in-process solo; s1 keys: 4/5 pass with routed beating solo, but algebra is suppressed outright 0.125 vs 0.958 — the cross-expert interference the s0 decouple map exposed, now in the routed row; fails the literal memorization-inflated prereg reference at both seeds); P5a fails, driven by arith_multi (0.04 own-drop) and algebra (0.00) — baseconv/fractions are locally surgical; usage purity ~0.13–0.31; P5d fails and confirms the exp011 blend-escape law at a second family (all five fail the ≥3× damping target in both runs: primary ratios 0.86–1.46, control 1.02–1.60); P5c resolved to a grounded negative by the e014c retry — first component 1.0 for every config, both two-step composites ~0.0 for every config: dispatch doesn't buy composition; the e014d CoT re-judge then sharpened it: the bare trunk composes at ceiling in CoT (add-then-percent 1.0) and the full collective degrades it to 0.21/0.0 — the blend-escaping experts measurably trample multi-step reasoning (P5d's amplitude finding and this task damage are one phenomenon); the direct-algebra control reproduces P5a/P5b/P5d and the original vacuous P5c (the e014c/d retries ran on primary keys only) — robust to the format confound
015	—	unassigned	numbering note: no v35 exp015 exists — the label was left unassigned to avoid collision with a parallel line's exp015 references
016	`exp004_caption/halluc.jsonl`	complete (s0+s1)	hallucination-reduction law HOLDS at both seeds: caption precision 0.356 → 0.694/0.705 while the invented-attribute rate falls 0.200 → 0.136/0.101 — conditioning reduces invention, not just adds coverage
017	`exp017_matrix/`	complete (s0)	full-matrix completion (single-anchor in-dispatch rows): caption survives at ~gauge under every specialist mask (0.36–0.37 at n=12; 0.389–0.406 vs gauge 0.4076 at n=48, term 1.0); own-task validity preserved; only_depth reproduces the depth-content dilution (score 0.09 at n=12, 0.123 at n=48, at validity ~1.0)
018	`exp018_retrospill/`	complete (retroactive, exp005 ckpts)	solo-specialist spillover matrix (the sweep that should have run first): every always-on specialist tramples every other task — depth destroys captioning (F1 0.003 at n=12; 0.0014 at the n=48 densification, terminate 0.0); paired with exp017, the strongest direct evidence that the dispatch contains specialize-regime anchors. 2026-07-18 audit amendments applied across exp005–011
019	`exp019_spangate/`	complete (s0, demo scale, n=12)	turn-granular span gating is functional: from the frozen floor (bbox 0.0, prose 0.42), only_bbox on the detect turn recovers 0.50/0.50 (= exp017's static only_bbox row; 0.5-valid × 1.0-conditional-score = 0.50 unconditional) and only_caption delivers full prose; all-on modestly better on detection — span control buys isolation/controllability, not accuracy. Two blind first runs disclosed in the package (input_ids-only generation: a silent VLM failure mode — plausible output with no image)
020	`exp020_relearn/`	complete (s0, 450 steps)	relearnability: `new` re-initialized inside the trained collective reconverges to its marginal role within ~150 steps (usage ~0.04, orig 0.03), caption recovers 0.546→0.66, termination 1.0 throughout — the derailment never reappears during regrowth. The passenger role is an attractor; the controller arc closes from a third side
021	`exp021_s1battery/`	complete (s1 battery + s0 comparison cell, n=12)	the 2-seed behavioral comparison for the central artifact: the caption and depth dissociations replicate (depth's pure content kill at intact validity — 1.0 valid / 0.0 score — now 2-seed), fixation redundancy replicates, all-on levels match (caption 0.727 vs 0.743); bbox's decouple is partial at s1 (0.91→0.73, cross-carried); the 07-19 s0 comparison cell retracted the seed-inversion claim — off_new damage is near-identical at both seeds: `new` is a core-task contributor AND a composite derailer at both seeds (two instruments, two truths; bbox is the only seed-dependent cell)

Plan-completion accounting (2026-07-18)

Against the campaign plan as written: every named experiment, instrument, gate, and deliverable ran and shipped, with three recorded exceptions and one recovered item. (1) OCR and classification specialists — band-excluded with rationale (OCR zero-shot floor 0.005; classification was plan-conditional and fell to the same band doctrine): the exp006 trap generalized — training against a cell with nothing learnable (ceiling or floor) buys only interference. (2) The big-JSON anchor — dropped by operator ruling mid-campaign in favor of the math-night program (exp012–014). (3) The span-limited lesion (Day-3, unconditional in the plan) — not run: a controller-signature probe made moot when exp010 resolved the controller hypothesis NOT SUPPORTED with a mechanism classification; exp019's span-gating demo covers the shared infrastructure from the control side. Named here rather than silently folded. Recovered: the relearnability test, initially declined for the same controller-moot rationale, ran in the restored window as exp020 — the re-grown new reconverges to its marginal role within ~150 steps with no derailment: the passenger role is an attractor.

Late-built instruments (vlmbench cross-judge, hallucination decomposition, the spillover sweep) were retro-applied to stored checkpoints and a full post-sweep audit amended six packages — see the dated amendment blocks in exp005–011.

Zero-shot baselines (Night 0: 15 vision tasks × 200 images, bf16, greedy)

The instruct model's fingerprint before any adapter — the baseline every conditioning claim is read against. schema_valid = parseable + schema-conformant JSON; score = task-specific deterministic judge.

task	valid	score	task	valid	score
image_classification	0.965	0.000	camera_rotational_offset	0.945	0.487
bbox_grounding	0.000	0.000	geometric_3d_object_id	1.000	0.000
ocr_text	0.005	0.005	semantic_association	1.000	0.031
segmentation	0.410	0.000	structural_spatial_awareness	1.000	0.000
outline_association	0.000	0.000	style_structural_awareness	1.000	0.400
subject_fixation	0.000	0.000	data_type_differentiation	0.675	0.090
depth_analysis	1.000	0.083	data_type_utilization	0.080	0.080
vit_accuracy_to_prompt	0.540	0.435

The always-on tax ladder (the series' central measurement)

Wikitext ppl@512 cost of adapter machinery on this substrate, all under identical gauges:

configuration	tax
one always-on register stack (exp002, 2 seeds)	+9.2 / +9.9
one ungated math stack (exp006 — fires at ~90% on-task amplitude on neutral text)	+91.6
five anchors under damped aleph dispatch (exp007)	+11.0
one always-on multi-task monolith stack, same four image tasks (exp009)	+3.66

Five anchors ride at roughly the cost of the exp002 register stack and ~8× under the measured ungated worst case — but ~3× the exp009 monolith, which carries the same four tasks always-on for +3.66 (candidate, s0). Always-on is therefore not intrinsically expensive: the tax tracks what the stack learned (math +91.6; text registers +9.2/+9.9; multi-task image monolith +3.66), and most of the collective's +11.0 belongs to the dispatch machinery itself. What that premium buys is caption fidelity (0.743 vs the monolith's 0.588) and composite composition (exp010: spec_only 0.417/0.167 vs monolith 0.0/0.167) — not ppl economy. Read the ladder's limits honestly: the first three rows are different machinery on different training data, sharing only the wikitext gauge; only the exp009 row is like-for-like with exp007 (same tasks, same data mix). The per-anchor decomposition landed in exp011 (P4): specialists damped 5–11× on neutral text, the blend-dominant caption anchor effectively undamped — no measured sub-additivity is claimed. [2026-07-18] The ladder prices always-on machinery in wikitext ppl only; exp018 adds the task-space column — always-on single-task stacks cost near-total destruction of the other three tasks, and the dispatch (exp017) is what buys that back. "Always-on is not intrinsically expensive" is scoped to multi-task co-trained stacks (see exp009's amendment).

Laws under test (running tally)

law (2.5-line status)	3.5 status
relay retrofit works on a frozen trunk	candidate, s0 (ppl 24.4→14.3 @512; remains s0)
gate-growth = the trunk opting in	not replicated — gates shrink while ppl improves (1 seed, flagged)
adapter deltas safe on recurrent state (new)	candidate, s0 — no runaway within 9k (ratio +2.3→+4.0%, still rising at window edge)
register capture → multi-task resolution (2 seeds)	validated w/ caveats (2 seeds; seed lottery; strong baseline)
specialists as seed-robustness	rationale supported (exp002 seed lottery, 2 seeds); specialist seed-robustness itself untested on 3.5 (exp005 is s0-only)
no-correctness-tax	on task correctness: holds at ~solo levels for 3 of 4 anchors at A=5 (exp008 all-on; candidate, s0, n=12/cell) with a −0.14 depth content dip flagged. Separately, the general-LM (ppl) side is quantified by the tax ladder above — damping bounds but does not eliminate (+11.0 at A=5)
termination-as-sample-semantics (2 seeds)	directional candidate at 12× length (rows > packed, n=8 s0; instruct prior softens the dichotomy)
two-regime dispatch / probe prediction (3-for-3)	first multimodal confirmation (candidate, s0, n=12/cell): blend-regime anchor (fixation, P1 sep 0.279) dissolved into redundancy; domain-distance anchors specialize-and-carry (3 double dissociations, exp008). Threshold calibration does NOT transfer to shared-image registers (all pairs ≤0.465) — the law holds directionally
surgical decoupling (damped one-off masks)	replicates multimodally (candidate, s0, n=12/cell) — exp008's table
dilution/starvation boundary (w/ safeguards)	zero alarms at A=5 across 5,000 steps (usage_ppl 2.3–4.1); depth shows a −0.14 dilution dip vs solo (n=12, flagged)
denominator damping at scale	bounded, not economical: +11.0 for five anchors is sub-additive only relative to the series' reference stacks; the exp009 always-on monolith carries the same four tasks at +3.66 (candidate, s0), so the dispatch machinery itself owns most of the collective's tax; per-anchor re-measure (P4) landed in exp011: specialists damped 5–11×, blend-dominant caption undamped
controller hypothesis (P3, falsifiable)	not supported (exp010; candidate, s0, n=12/cell): no-absorber screen passed (exp008), but all_on scores 0.0 on composites and spec_only (`new` removed) rescues to 0.417/0.167 — the preregistered ≥+0.15 threshold missed by −0.417/−0.167. Composition emerges from the dispatched specialists, not from a trained controller (monolith: 0.0/0.167, below spec_only). Magnitude/mechanism classification pending the raw-output diag
toggle gates bit-exact	confirmed (exp011, candidate s0) — all-off reproduces the frozen trunk at max
P4 damping decomposition (preregistered)	landed with a finding (exp011): specialists damped 5–11× on neutral text, but the blend-dominant caption anchor rides at ~solo-ungated amplitude (0.100 vs the 0.102 reference) — damping protects against specialize-regime anchors; blend-regime anchors escape it. Decomposes the +11.0 tax
THE REGIME LAW at family scale (the series' headline)	measured on both sides (e017/e018 + exp014, 2026-07-18): always-on solo specialists are mutually destructive (depth stack reduces captioning to F1 0.003/term 0.0); the aleph dispatch rescues specialize-regime anchors almost completely (caption back to ≈gauge under every single-anchor mask, own-task validity preserved) and fails blend-regime anchors (2-seed on the failure verdicts; the one positive, tax-free stacking, is seed-fragile where interference lands) (math experts: no damping — ratios 0.86–1.6 vs the ≥3× target — and active CoT trampling 1.0→0.21). The one reproduced rescue cost: in-dispatch depth content dilutes (0.826→0.123 at n=48). Regime decides whether dispatch helps — readable in advance from sign-code separation for the image family (P1 probes, directional) and from solo-training provenance for the math experts. exp011 P4's amplitude decomposition predicts which anchors the dispatch contains, anchor for anchor
blend-escape law generalizes (caption → math)	confirmed at a second, independently built family (exp014 + control, candidate s0): all five solo-trained math experts enter the dispatch as blend-regime anchors and fail the ≥3× damping target in both runs (on-domain/neutral amplitude ratios: primary 0.86–1.46, control 1.02–1.60) — the dispatch stacks them accuracy-tax-free on band cells (P5b) but does not silence them off-domain
AND-token doctrine, question-space extension (new)	minted (exp013): the answer-diversity guard does not protect against question-space exhaustion — sequences t2 (480 distinct questions) and baseconv t1 (248) passed the answer guard yet were memorized whole by 800 training draws/tier; the held-out re-judge found literally no unseen question to test. A companion guard `guard_train_space` (space ≥ 3× draws) ships in `mathgen.py`, adopted as a precondition for future expert selection (no training has yet run behind it)
caption deliverable seed stability (held-out cross-judge)	validated (exp004b amendment): frozen 0.408 → s0 0.706 / s1 0.704 token-F1 on 50 held-out images, \|s0−s1\| = 0.0019, terminate/nonempty 1.0 across all three configs — the +0.30 delta replicates across seeds on held-out data, independently re-judged
predictability principle (LM loss teaches what target-token predictability demands)	validated cross-exp on 3.5, assembled retrospectively (2026-07-18): (i) exp005 fixation — targets predictable without vision buy format, not skill ("format anchor": localization already 0.94 when format succeeded); (ii) exp016 — caption tokens unpredictable without the image (garments/colors) force grounding: precision 0.356→0.70, invention halved; (iii) exp013 algebra-steps — chained derivation tokens force computation (held-out ceiling). The 2.5-line principle transfers; no dedicated bed was run — the plan's "checkable image constraints" instrument existed implicitly across these three
OCR specialist (plan Day-2 item)	band-excluded, documented (2026-07-18): zero-shot OCR is a floor (0.005 valid / 0.005 score, Night-0 fingerprint) — below the 0.10–0.60 learnable band the exp012 doctrine later formalized. Training against it is the exp006 trap generalized (ceiling or floor: nothing learnable, so a degenerate anchor buying only interference). Excluded from the specialist sweep with this rationale on the record rather than silently skipped
caption deliverable artifact + loader gates (e004c/d)	both PASS: the hub checkpoint re-downloaded fresh reproduces F1 = 0.706 exactly (e004c), and the standalone `relay_caption.py` loader on a plain-transformers trunk (no project substrate) reproduces 0.706 exactly (e004d) — weights and shipped code both validated. The complete bundle (card + loader + both seeds) lives at `caption_deliverable/` and as the dedicated repo AbstractPhil/qwen3.5-0.8b-relay-caption

Math night (exp012–014): map the weakness, train the experts, dispatch the basins

One overnight arc (replacing a heavier data-generation plan). Its discipline was set by exp006's failure (never train against a ceiling): first a weakness scan with an internal control that must reproduce the known ceiling (it did, 1.0), then experts only inside the learnable band, then the basins collective under preregistered verdicts. Three durable findings:

Where 0.8B math weakness lives: mostly in format (CoT rescues many direct floors outright — the capability exists but cannot compress into a direct answer) — and where it is genuine (multi-digit products, fraction arithmetic), an expert either lifts it cleanly (algebra to 0.83 in-sample and 1.00 held-out; fractions to 0.83 in-sample, 0.70 held-out) or hits a wall (arith t3: +1/24, noise).
The question-space law: two "striking" expert gains (sequences 0.21→1.00) dissolved under the held-out re-judge — their generators' question spaces were smaller than the training draws, so the experts memorized the space. The answer-diversity guard passed these cells; the new question-space guard would not have.
Blend-escape transfers, and composition is a grounded negative: the aleph dispatch stacks five frozen math experts with no band-cell accuracy tax (P5b) — but grants neither surgical independence (P5a) nor off-domain damping (P5d): solo-trained always-on experts are blend-regime anchors, exactly as exp011 found for caption, now in a second, independently built expert family. And composition (P5c, e014c retry): with the first component provably at ceiling for every config, both two-step composites floor for trunk, solo experts, and collective alike — dispatch does not buy composition. Then the CoT re-judge (e014d) sharpened the negative into a mechanism: the bare trunk composes at ceiling with visible steps (add-then-percent 1.0), and the full collective actively degrades it (0.21 / 0.0 — the worst config on both cells). The blend-escaping experts don't merely evade damping; they trample multi-step reasoning the trunk handles natively — P5d's amplitude finding and this task damage are one phenomenon seen from two sides. A collective that wants quiet, composing basins needs selectivity trained in, not imposed by the denominator.

Every number above was adversarially audited before ship (exp012: two independent audit passes; exp013/014: one shared adversarial pass plus a follow-up on this summary; all corrections are disclosed in-package, including a retracted overclaim of our own).

Port surfaces (what did NOT transfer from 2.5)

Tool-call render format: Qwen3.5's template emits XML-parameter blocks (<function=…><parameter=…>), not <tool_call>{json}</tool_call>. The 2.5-line parser silently zeroes every valid output (see exp002's amendment).
Batch-1 decode is a ~32 tok/s ceiling on the hybrid trunk with the naive torch DeltaNet fallback (one CPU core pinned, GPU idle). Batched generation recovers 10–13×; all judges in this series batch.
Fused linear-attention kernels (fla) are three-way blocked on torch 2.4.1 (fla 0.5.x needs triton ≥3.3; torch 2.4 pins 3.0; fla 0.1–0.2 needs torch ≥2.5 APIs). This series runs the torch fallback uniformly — slower, but numerically consistent across every run.
Judge windows must cover the trained length distribution: a 2,048-token generation cap scored two healthy 3–4.5k-token emitters at exactly 0.0 (exp003's amendment) — the judge measured its own ceiling.
M-RoPE needs mm_token_type_ids in hand-built training batches, and native-resolution images can exceed a whole prefix budget on their own (exp004's gates; both handled inside substrate/vlm_collator.py).