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

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

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

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

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.

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

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

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)

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:

1. 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).
2. 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.
3. 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)

1. 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).
2. 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.
3. 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.
4. 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.
5. 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).

All runs on a single RTX 6000 Ada (48GB), wall-clock ledgered.