Burndown publication set: production README (final statuses, laws tally, accounting), article_ft3d (the conditioning law), REPO_MAP (both lines)

README.md

@@ -1,58 +1,76 @@
 ---
 license: mit
+base_model:
+- stable-diffusion-v1-5/stable-diffusion-v1-5
+- AbstractPhil/sd15-flow-lune-json-prompt
+- circlestone-labs/Anima
 tags:
 - diffusion
 - adapters
 - aleph
 - geolip
+- research-record
 ---
 
-# geolip-aleph-diffusion
-
-Testable **aleph adapters** for the geolip diffusion variants — the certified
-components of the geolip-aleph-qwen text line (frozen-trunk relay adapters with
-closed-form aleph addressing, zero-init out weight+bias, near-zero gates,
-bit-exact toggle contract) ported to diffusion trunks. Runner-2 line of the
-geolip program.
+# geolip-aleph-diffusion — the diffusion aleph campaign (runner 2)
+
+**This is the research record.** If you're here for something usable:
+
+| you want | go to |
+|---|---|
+| production adapters (safetensors, curated) | [aleph-diffusion-adapters](https://huggingface.co/AbstractPhil/aleph-diffusion-adapters) |
+| the framework (train/attach/toggle/detach + diagnostics) | [amoe-lora](https://huggingface.co/AbstractPhil/amoe-lora) (`amoe.diffusion`, 0.2) |
+| the multi-GPU trainer | [diffusion-pipe fork](https://github.com/AbstractEyes/diffusion-pipe) (`aleph_relay = true`) |
+| the article | [article_ft3d.md](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/blob/main/article_ft3d.md) |
+| the repo map of the whole program | [REPO_MAP.md](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/blob/main/REPO_MAP.md) |
+
+**The scientific headline — the conditioning law (2-seed):** structural
+(segmentation-blob) supervision pays where x0 recovery is LINEAR and is
+inert where it is ill-conditioned. The identical foreground-LP-x0
+coupling moved its gauge **−5.9%/−3.7% (two seeds) on the flow substrate
+vs +0.03%/−1.0% on eps — a ~125–200× effect ratio** — because flow
+recovers `x0 = x_t − σ·v` exactly at every sigma while eps divides by a
+vanishing `sqrt(alpha_bar)` precisely in the supervised band
+([exp012](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp012_blob_supervised) /
+[exp013](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp013_blob_flow)).
+A 3-point dose curve fixes **λ≈1** as the operating point. Two companion
+results: the relay machinery certified **relay-favorable 3-for-3
+substrate classes** (Lune flow / SD15-core eps / Anima DiT flow — matched
+LoRA controls landed at-or-below frozen every time), and the **multiband
+mechanism** manufactures true sigma-band specialists (band lesions
+surgical 3/3, own-band damage 50–200× cross-band, both seeds) whose
+payer is role-aligned structure (multiband beats the matched monolith
+~10% on HIGH-band foreground structure, 2-seed) — invisible to every
+aggregate-eps comparison.
+
+Every result reads against the exp000 baseline wall; paired designs with
+derangement controls; judges self-test before scoring; 1-seed results
+ship as **candidates**. Amendments and corrections ship as dated blocks
+inside the affected package — they are not blemishes; they are the
+method.
 
 ## Substrates
 
 | variant | trunk | trainer path |
 |---|---|---|
 | **SD1.5 core (epred) — primary** | `stable-diffusion-v1-5/stable-diffusion-v1-5`, epsilon objective, natural 77-token prompting | custom beds (this repo, `substrate/`) |
-| **SDXL core (epred) — primary** | `stabilityai/stable-diffusion-xl-base-1.0` | after SD1.5-core certifies; plus the geolip-sdxl-aleph read-only capacity battery |
-| SD15-Lune rectified flow (undertrained exemplar; cheap flow testbed) | `AbstractPhil/sd15-flow-lune-json-prompt` ckpt-2500 (+ base lune, json-vit) | custom beds (this repo, `substrate/`) |
-| Anima / Cosmos-Predict2 2B DiT | `circlestone-labs/Anima` (NC weights — all derived checkpoints NC) | `AbstractEyes/diffusion-pipe` branch `feat/aleph-adapter` |
-| Krea 2 Turbo — expansion | `krea/Krea-2-Turbo` (distilled from `krea/Krea-2-Raw`, the trainable base; Krea license — check before derived ships) | fork `krea2` path; roster entry, beds TBD |
-
-**Prompting congruency:** judged work carries BOTH the json-225 chunked
-encoding and standard 77-token plain-English prompting (congruency to the
-natural paradigm).
-
-Sampler of record (Lune family): rectified flow, `v = noise − x0`, Euler on the
-**SHIFT=2.5-warped** sigma grid `σ = 2.5u/(1+1.5u)`, `VAE_SCALE 0.18215`. Text
-conditioning is the trainer's 225-token 3-chunk CLIP encoding (`encode_clip_225`,
-227 positions) — never a bare 77-token encode.
-
-## Standing rules of the line
-
-- Adapter dtype **matches the trunk dtype** (fp32 adapters on a low-precision
-  trunk inject fp32 noise; judged gauges still run fp32).
-- Pure Adam, weight-decay 0 on adapter paths — never AdamW.
-- No comparative selectors (argmax/softmax routing/VQ) anywhere near the aleph;
-  the addresser is the closed-form sinh/cosh read.
-- All-adapters-off ≡ frozen trunk **bit-exact** (code-path-skip toggles), gated
-  before and after every training run.
-- Every result reads against the exp000 zero-shot baseline wall below; paired
-  designs with derangement (no-fixed-point) controls; judges self-test on known
-  answers before scoring; 1-seed results ship as **candidates**.
+| SD15-Lune rectified flow (undertrained exemplar; the cheap flow testbed) | `AbstractPhil/sd15-flow-lune-json-prompt` ckpt-2500 (+ base lune, json-vit) | custom beds (this repo, `substrate/`) |
+| Anima / Cosmos-Predict2 2B DiT | `circlestone-labs/Anima` (NC weights — all derived checkpoints NC) | `AbstractEyes/diffusion-pipe` (aleph baked into main) |
+| SDXL core (epred) | `stabilityai/stable-diffusion-xl-base-1.0` | fork sdxl aleph path landed; battery = next campaign (exp005 reserved) |
+| Krea 2 Turbo — expansion | `krea/Krea-2-Turbo` (distilled from `krea/Krea-2-Raw`; license check before derived ships) | fork `krea2` path; roster entry, beds TBD |
+
+Sampler of record (Lune family): rectified flow, `v = noise − x0`, Euler
+on the **SHIFT=2.5-warped** sigma grid `σ = 2.5u/(1+1.5u)`, VAE scale
+0.18215; text conditioning `encode_clip_225` (227 positions), never a
+bare 77-token encode. **Prompting congruency:** judged work carries BOTH
+json-225 and standard 77-token plain-English prompting. Judge floor:
+CLIP-L image-image cosine has a ~0.988 pure-noise floor — gaps live in a
+narrow band; the paired design is load-bearing.
 
 ## Baseline wall (exp000)
 
-CLIP-L image↔image cosine, original vs regenerated from the row's own
-`json_prompt` (`vit_json_prompt` for json-vit), **conditional vs shuffled-JSON
-(derangement)**, n=24 held-out rows of `synthetic-object-relations-json`,
-paired seeds, 30 steps, guidance 6.0, fp32 judge features:
+CLIP-L image↔image cosine, cond vs shuffled-JSON (derangement), n=24
+held-out rows, paired seeds, fp32 judge:
 
 | model | cond | shuffled | **cond − shuffled** |
 |---|---|---|---|
@@ -60,33 +78,73 @@ paired seeds, 30 steps, guidance 6.0, fp32 judge features:
 | json_ckpt2500 | 0.7536 | 0.5437 | **+0.2099** |
 | base_lune (18765) | 0.6347 | 0.5272 | **+0.1075** |
 
-Reading: the json-prompt finetunes carry ~2× the base UNet's scene information
-from structured-JSON conditioning. Caveats: n=24, single seed bank; CLIP-L
-image-image cosine has a high floor (~0.988 for pure-noise pairs on identical
-resolution) so gaps live in a narrow band — the paired design is load-bearing.
+Natural-paradigm wall (exp000b): format specialization is a **double
+dissociation** (base_lune nl77 +0.2043 vs json +0.1075; json_ckpt2500
+DROPS on nl77); the epred cores are strong natural carriers zero-shot
+(sdxl_core +0.2129 tops the NL wall; sd15_core +0.1809).
 
-## Experiments
+## Experiments (final statuses)
 
-| exp | question | status |
+| exp | question | verdict |
 |---|---|---|
-| exp000_baselines | zero-shot baseline wall (above) | **shipped** |
-| exp000b_natural | natural-paradigm wall: format specialization is a double dissociation; SDXL core tops the NL wall | **shipped (candidate, s0)** |
-| exp001_sd15_relay | relay-all16 vs matched LoRA-r32 vs frozen: **relay beats both; LoRA lands below frozen; post-train toggle bit-exact** | **shipped (candidate, s0)** |
-| exp002_sd15_addrcond | addr-cond 4-arm causal: **GUIDEPOST with a pulse** — inert beside full text (real≈deranged), but the address ALONE steers (+0.029); redesign = complementarity | **shipped (candidate, s0)** |
-| exp003_sigma_registers | register probe: **sigma is the ONLY live axis** — caption/type/cond all compress to ≤0; ordering prereg MISS with design consequences | **shipped (candidate, s0)** |
-| exp004_anima_relay | ANIMA DiT via the fork: **relay improves (0.123→0.115) while native LoRA degrades (0.138→0.142) — relay-favorable 3-for-3 substrates; full fork path certified; ckpt NC** | **shipped (candidate, s0)** |
-| exp005_sdxl_tree4a | the SDXL capacity battery (guidepost/scaffold/skeleton) | designed |
-| exp006_sd15core_relay | CORE EPRED certification: **relay beats frozen (−2.5%) and matched LoRA 2-for-2; grounding gained (+0.036) where LoRA traded it; gates GROW on the core** | **shipped (candidate, s0)** |
-| exp007_amoe_lora | AMoE v1 vs monolith falsifier: **falsifier fires — uniform pressure ⇒ flat usage, dispatch pays nothing (Tree-1b law on diffusion); machinery green; next = structural band assignment** | **shipped (candidate, s0)** |
-| exp008_multiband | band-ASSIGNED experts: **mechanism certifies — 3/3 surgical band lesions (50-200x), HIGH-noise band expert wins its band (label-corrected); uniform objective doesn't pay (roles enter exp009)** | **shipped (candidate, s0; s1 running)** |
-| exp009_bandroles | role objectives: **directional hit 4/4 but noise-adjacent — frequency reweighting too collinear; needs qualitatively different supervision + generation-side gauges (exp010)** | **shipped (2 seeds + rejudge)** |
-| exp010_controller | **StepGatedSampler ships**: controller lifts grounding +0.089 over frozen; monotonic lesion ladder; HIGH lesion 14x LP-dominant (coarse-to-fine confirmed in image space); eps-trained HIGH band concentrates (diversity = open training goal) | **shipped (candidate)** |
-| exp011a_fused_multiband | multiband on REAL fused data: **adapters pay 2-3x more; structural story replicates (3/3 surgical, monolith edge persists); blob targets built 100% after a schema lesson** | **shipped (candidate, s0; s1 running)** |
-| exp013_blob_flow | **CONDITIONING HYPOTHESIS CONFIRMED**: same blob coupling, ~200x the eps effect on the flow substrate (−5.9% vs +0.03%); blob supervision belongs on flow/v-pred trunks | **shipped (candidate, s0; s1 running)** |
-| exp014_te_dispatch | router-solidifier v1: **honest negative** — state keys already route by prompt; raw-flattened address kills routing (the known high-D disease geometry); null-design lesson → v2 | **shipped (candidate, s0; s1 running)** |
-| exp015_router_v2 | corrected router test: **address-as-key FALSIFIED** (routing excess 2.5e-06 over constant-key null; match advantage −0.0); open form = TE+AMOE joint training | **shipped (candidate, s0)** |
-
-Numbering: a number binds to a preregistered design at assignment and is never reassigned — gaps mean reserved-not-yet-run (exp005 = the SDXL Tree-4a battery, designed; exp004 closed 2026-07-19). Sub-letters (000b) extend a shipped experiment without disturbing its rows.
-
-Code for each experiment ships in its folder with `results.json` from the run
-ledger. Substrate modules in `substrate/`.
+| [exp000_baselines](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp000_baselines) | zero-shot baseline wall | **shipped** |
+| [exp000b_natural](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp000b_natural) | natural-paradigm wall | **shipped** (wall; format double dissociation) |
+| [exp001_sd15_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp001_sd15_relay) | relay vs matched LoRA vs frozen (Lune flow) | **2-seed vs frozen** — relay −0.85%, grounding preserved; LoRA BELOW frozen, grounding degraded; toggle bit-exact |
+| [exp002_sd15_addrcond](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp002_sd15_addrcond) | address into the cond stream | **honest negative w/ a pulse** (candidate s0) — inert beside full text; addr ALONE steers (+0.029); redesign = complementarity |
+| [exp003_sigma_registers](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp003_sigma_registers) | register probe | **final** — sigma is the ONLY live axis; caption registers negative everywhere |
+| [exp004_anima_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp004_anima_relay) | Anima DiT through the fork | **candidate s0** — relay improves (0.123→0.115) while native LoRA degrades (0.138→0.142); full fork path certified end-to-end; ckpt NC |
+| exp005_sdxl_tree4a | SDXL capacity battery | **reserved, designed, not run** (numbering law: gaps mean reserved) |
+| [exp006_sd15core_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp006_sd15core_relay) | CORE EPRED certification | **2-seed** — relay −2.5% vs frozen, beats matched LoRA; grounding +0.1809→+0.2172 where LoRA traded it away; gates GROW on the core (substrate-dependent dynamics) |
+| [exp007_amoe_lora](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp007_amoe_lora) | AMoE v1 vs monolith falsifier | **falsifier fires, 2-seed** — uniform pressure ⇒ flat usage; Tree-1b law reproduces on diffusion |
+| [exp008_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp008_multiband) | band-ASSIGNED experts | **mechanism certified 2-seed** — 3/3 surgical band lesions (50–200×); uniform objective doesn't pay (Law-1 again); label correction confessed + reshipped |
+| [exp009_bandroles](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp009_bandroles) | HP/LP role objectives | **directional-but-negligible, 2 seeds + rejudge** — frequency reweighting ~collinear with eps; instrument gap confessed, rejudged |
+| [exp010_controller](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp010_controller) | the step-gated controller + image-space battery | **shipped** — controller grounding +0.2135 vs frozen +0.1247; monotone lesion ladder; HIGH lesion 14× LP-dominated (coarse-to-fine in image space); eps-MSE confirmed gauge-blind |
+| [exp011_fused_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp011_fused_multiband) | multiband on REAL fused data | **2-seed** — real data pays 2–3×; structural story seed-robust; blob targets rebuilt 100% after a schema lesson |
+| [exp012_blob_supervised](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp012_blob_supervised) | blob supervision on eps + the gauge discovery | **2-seed** — blob-on-eps INERT (the conditioning control); the ROLE-ALIGNED GAUGE discovery: multiband beats monolith ~10% on HIGH-band fg structure (0.1115 vs 0.1234; frozen 0.1555) |
+| [exp013_blob_flow](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp013_blob_flow) | the same coupling on FLOW | **CONDITIONING LAW 2-seed + 3-point dose curve** — −5.9%/−3.7% vs eps's +0.03%/−1.0% (~125–200×); λ≈1 operating point; undertrained lune pays adapters −9.1% (headroom scales inversely with trunk training) |
+| [exp014_te_dispatch](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch) | text/address in the dispatch key | **honest negative w/ teeth, 2-seed** — state keys ALREADY route by prompt; raw-flattened address KILLS routing (disease geometry); null-design lesson |
+| [exp015_router_v2](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp015_router_v2) | corrected address-as-key test | **FALSIFIED 2-seed** — routing excess 2.5e-06 over the repeated-key null; match advantage ~0; open form = TE+AMoE joint training |
+
+Numbering law: a number binds to a preregistered design at assignment and
+is never reassigned — gaps mean reserved-not-yet-run. Sub-letters (000b)
+extend a shipped experiment without disturbing its rows. Every package
+carries its code, `results.json`, and ledger rows; adapter weights ship
+as `.pt` + bitwise-verified `.safetensors` companions
+(`amoe.diffusion.anchor` format).
+
+## Laws under test (text line → diffusion status)
+
+| law | diffusion verdict |
+|---|---|
+| toggle law (all-off ≡ frozen bit-exact) | **HELD in every experiment**, post-train included |
+| zero-init W **and** B (bias-leak law) | encoded; P-INIT bit-exact at every attach |
+| Law-1: architecture alone doesn't pay; differentiated pressure does | **reproduced twice** (exp007 falsifier, exp008 uniform-objective monolith edge) |
+| Law-2: honest negatives fire before scale spend | **fired** (exp002 addr-in-context; exp009 role reweighting) |
+| gate-growth (2.5-line) | **substrate-DEPENDENT** — core relay gates GROW 0.047→0.070 (trunk opts in), lune gates shrink; neither in the LM band. Scope finding |
+| comparative routing near the aleph | **falsified on diffusion 2-seed** (exp007/014/15); structural banding is the certified alternative |
+| conditioning law (NEW, diffusion-native) | **2-seed + dose curve** — structural supervision needs linear x0 recovery |
+| dtype law (NEW) | adapter dtype = DECLARED trunk dtype; violation caught live by the R0b gate (meta-tensor sniff) and fixed in the fork |
+
+## Plan-completion accounting
+
+Receipts-forward, per house custom: a **5.7h idle gap** (2026-07-19,
+turn ended on an intention without a launch — confessed, recovered, the
+launch-first law re-earned program-wide); the **exp008 label correction**
+(band labels inverted in the first shipped README — numbers right,
+labels wrong; corrected + reshipped same hour); the **exp009 instrument
+gap** (reference arms lacked role gauges as first run — confessed,
+rejudged); exp015-s1 ran on ~3h of **disclosed reserve** (no-downtime
+directive; pod paid regardless). Campaign totals: 47 queue jobs, ~61h
+GPU ledgered at ~97% utilization inside the window, 16 experiment
+packages + substrate, every claim 2-seed or bounded. The complete
+stdout record (27 run logs + `pod2_ledger.jsonl`) is archived in the
+program's research memory; per-package ledgers ship here.
+
+## Open questions the record points at
+
+v-pred same-trunk arm (the sharpest conditioning-law follow-up);
+trained-diversity pressure for the HIGH band (exp010's G3 flip);
+TE+AMoE joint routing (the open form after exp015); exp016 assembly
+(relay + multiband + controller composed); SDXL battery (exp005,
+reserved); Krea-2 turbo beds; qualitatively different band supervision
+(segmentation/semantics beyond frequency reweighting).

REPO_MAP.md

@@ -0,0 +1,26 @@
+# Repo map — the geolip-aleph program
+
+## The Qwen (language) line
+
+| repo | what it is | what you take from it | maturity |
+|---|---|---|---|
+| [geolip-aleph-qwen-3.5-0.8b-instruct](https://huggingface.co/AbstractPhil/geolip-aleph-qwen-3.5-0.8b-instruct) | **the research record** (the ft3 campaign): 19 experiment packages, raw ledgers, audits, retractions, `article_ft3.md`, `campaign_ledger.jsonl` | the evidence — every claim's receipts; all checkpoints under `*/ckpt/` | complete, frozen at campaign close |
+| [qwen3.5-0.8b-relay-caption](https://huggingface.co/AbstractPhil/qwen3.5-0.8b-relay-caption) | **production caption adapter** for Qwen/Qwen3.5-0.8B | `relay_caption.py` + both seed checkpoints; validated end-to-end (artifact + loader gates PASS, off-pod quickstart) | usable; know the e004e trade-off (detach for structured vision tasks) |
+| [geolip-aleph-qwen](https://huggingface.co/AbstractPhil/geolip-aleph-qwen) | the **Qwen2.5-0.5B predecessor line** (exp001–014 there) + `article_ft2.md` | the two-regime dispatch law's first certification; the ft2 article | complete (prior campaign) |
+| [qwen-deepfashion-fused](https://huggingface.co/datasets/AbstractPhil/qwen-deepfashion-fused) | the fused image dataset (captions + per-task JSON GT + entity polygons) | training/eval data for the caption line AND the diffusion line's fused-data + blob targets | stable |
+
+## The diffusion line (runner 2)
+
+| repo | what it is | what you take from it | maturity |
+|---|---|---|---|
+| [geolip-aleph-diffusion](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion) | **the research record** (the 3-D campaign): 16 experiment packages + substrate, 2-seed program, `article_ft3d.md`, all adapter weights (.pt + verified .safetensors) | the evidence — the conditioning law, relay 3-for-3, the multiband mechanism, the routing falsifications | complete, frozen at campaign close |
+| [aleph-diffusion-adapters](https://huggingface.co/AbstractPhil/aleph-diffusion-adapters) | **production adapters** (curated safetensors: sd15 core relay + multiband, lune-flow blob stack, anima relay) | attachable artifacts with provenance metadata; the quickstart | usable at research grade; per-row evidence status on the card; anima NC |
+| [amoe-lora](https://huggingface.co/AbstractPhil/amoe-lora) | **the framework**, now both substrate families: LM verbs + `amoe.diffusion` (relay/multiband/StepGatedSampler, safetensors I/O, `amoe-convert`) | the `amoe` package (pip -e .[diffusion]); invariant-tested incl. dtype law; every campaign checkpoint format loads | 0.2 working-state; train reference-grade; DDP untested >1 GPU; FSDP non-goal; ComfyUI node next |
+| [diffusion-pipe fork](https://github.com/AbstractEyes/diffusion-pipe) | **the multi-GPU trainer** with aleph baked into main (cosmos/anima relay+multiband, SDXL relay; amoe-format saves; pure-Adam guard) | `aleph_relay = true` in any supported TOML; `examples/aleph_*.toml` | proven end-to-end on Anima (exp004); SDXL path landed, site count pins at first run; multiband stages>1 deferred |
+
+Articles: [ft1](https://huggingface.co/blog/AbstractPhil/aleph-autoregressive-differentiation-ft1)
+(byte-level origins) → [ft2](https://huggingface.co/AbstractPhil/geolip-aleph-qwen/blob/main/article_ft2.md)
+(Qwen2.5: two-regime law) → [ft3](https://huggingface.co/AbstractPhil/geolip-aleph-qwen-3.5-0.8b-instruct/blob/main/article_ft3.md)
+(Qwen3.5 VLM: the regime law at family scale) →
+[ft3d](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/blob/main/article_ft3d.md)
+(diffusion: the conditioning law).

article_ft3d.md

@@ -0,0 +1,312 @@
+# Aleph Differentiation, Part 3-D (Diffusion): The Conditioning Law
+
+*The aleph adapter program leaves language: relays on frozen denoisers,
+band experts on the sigma axis, a step-gated controller — and the law
+that decides where structural supervision can pay at all.*
+
+This is the diffusion branch of the aleph adapter program.
+[Part 1](https://huggingface.co/blog/AbstractPhil/aleph-autoregressive-differentiation-ft1)
+established the addressing architecture on byte-level models.
+[Part 2](https://huggingface.co/AbstractPhil/geolip-aleph-qwen/blob/main/article_ft2.md)
+certified the two-regime dispatch law on Qwen2.5-0.5B.
+[Part 3](https://huggingface.co/AbstractPhil/geolip-aleph-qwen-3.5-0.8b-instruct/blob/main/article_ft3.md)
+survived a second substrate and sharpened the law into a deployment
+instrument. Part 3-D asks a harder transfer question: the machinery was
+built for autoregressive token streams — **does any of it survive a
+generative model with no token axis, a sigma axis instead, and three
+different parameterizations of the same denoising problem?** The
+answer, after 16 experiment packages, a 2-seed program, and one
+four-day rented GPU: the *machinery* transfers almost embarrassingly
+well, the *routing* does not transfer at all, and the sigma axis gives
+back something the language line never had — a law that predicts, from
+the parameterization alone, whether structural supervision will pay
+before you spend a single GPU-hour on it.
+
+Everything below links to a self-contained package in the
+[research record](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion)
+with raw ledgers and self-asserting results files. The production
+artifacts are the
+[aleph-diffusion-adapters](https://huggingface.co/AbstractPhil/aleph-diffusion-adapters)
+repo (curated safetensors) and the
+[amoe-lora](https://huggingface.co/AbstractPhil/amoe-lora) framework,
+whose 0.2 release ships everything in this article as five verbs
+(`amoe.diffusion`: train / attach / toggle / detach, plus the
+controller); the multi-GPU trainer is the
+[diffusion-pipe fork](https://github.com/AbstractEyes/diffusion-pipe)
+with the adapters baked in.
+
+## The substrate is not a bystander
+
+The language line learned this lesson once; diffusion teaches it three
+times simultaneously. The campaign ran one adapter architecture against
+three substrate classes: **SD15-Lune** (an undertrained rectified-flow
+finetune — the cheap flow testbed), the **stock SD1.5 core** (epsilon
+objective, the "most powerful element" of the family and the primary
+target), and the **Anima/Cosmos-Predict2 2B DiT** (bf16, flow, trained
+through the DeepSpeed pipeline fork rather than our own beds). Three
+parameterizations, two architectures, two dtypes, two trainers.
+
+Two substrate findings frame everything else. First, **gate dynamics
+are substrate-dependent**: on the stock core the relay gates *grow*
+(0.047→0.070 — the trunk opts in, the Part-2 mechanism), on the
+undertrained Lune they *shrink*, and neither lands in the LM band. The
+2.5-line gate-growth law survives as a mechanism, not as a constant.
+Second, the dtype of the adapter is not a detail: attaching fp32
+relays to a bf16 trunk spins fp32 noise into the environment, and the
+integration gate caught our own code doing exactly that — sniffing a
+first parameter that was still a meta tensor and reading fp32 off it.
+The **dtype law** (adapters attach at the *declared* trunk dtype,
+never a sniffed one; judged gauges stay fp32) is now encoded in the
+framework and the fork both.
+
+The judging floor also deserves its sentence: CLIP-L image-image
+cosine has a ~0.988 floor on pure-noise pairs, so every grounding
+number below lives in a narrow band and the paired
+derangement-controlled design is load-bearing, not decorative. The
+zero-shot baseline wall
+([exp000](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp000_baselines),
+[exp000b](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp000b_natural))
+put numbers under every later claim — and produced its own finding on
+the way: conditioning-format specialization is a **double
+dissociation** (the base model doubles its grounding on plain-English
+prompts exactly where the JSON-finetuned variant drops), and the stock
+epred cores are strong natural carriers zero-shot.
+
+## The deliverable first: the relay certifies 3-for-3
+
+[exp001](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp001_sd15_relay)
+attached the certified 16-slot relay (zero-init output weight *and*
+bias; gate −3.0; bit-exact code-path-skip toggle) to all 16
+cross-attention blocks of the frozen Lune flow trunk, against a
+parameter-matched LoRA-r32 control and the frozen baseline. Paired
+val flow-MSE: relay **−0.85%** vs frozen, with round-trip grounding
+preserved (+0.2052 vs the wall's +0.2099) — while the matched LoRA
+landed *below frozen* (+0.85% worse) and degraded grounding outright
+(+0.2099 → +0.1694). On the primary substrate
+([exp006](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp006_sd15core_relay),
+stock SD1.5, epsilon objective, 2-seed) the pattern repeats stronger:
+relay −2.5% vs frozen, LoRA behind it, and the relay *gains* grounding
+over the zero-shot core (+0.1809 → +0.2172) exactly where LoRA trades
+it away (+0.1596). And on the third substrate class
+([exp004](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp004_anima_relay),
+the Anima DiT trained end-to-end through the diffusion-pipe fork,
+bf16, 28 blocks), the relay improves its loss (0.123→0.115) while the
+fork's own native LoRA control *degrades* over training (0.138→0.142)
+— the exp001 destabilization pattern on a third architecture.
+
+**Relay-favorable 3-for-3 substrate classes, with matched controls
+every time.** The toggle law held bit-exact post-train in every one of
+these experiments. That contract is why the production story is
+"attach, toggle, detach" and not "merge and hope": the deliverable
+adapters in
+[aleph-diffusion-adapters](https://huggingface.co/AbstractPhil/aleph-diffusion-adapters)
+are removable, verifiably, to the bit.
+
+## The sigma axis: bands, lesions, and a controller
+
+The language line's registers had no equivalent of noise level. The
+register probe
+([exp003](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp003_sigma_registers))
+made the reorientation unmistakable: caption-content registers
+separate *negatively* at every band and site (image content dominates
+the code surface — multimodal compression in the extreme), and
+**sigma is the only live register axis**, strongest at the deepest
+site. So the mixture question became: assign experts *structurally*
+along sigma instead of routing them comparatively.
+
+[exp008](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp008_multiband)
+built the mechanism: three band experts per site (LOW/MID/HIGH noise),
+combined by smooth cosine crossfade windows — positional gating, dense
+and differentiable, no selection event anywhere. The certification is
+the lesion battery: **disabling a band's experts damages that band's
+validation 50–200× more than any other band's, 3-for-3 bands, both
+seeds.** Band assignment manufactures true specialists. What it does
+*not* do by itself is pay on the aggregate objective — the matched
+monolith keeps a small edge under uniform pressure at both seeds,
+which is the language line's Law-1 (architecture alone doesn't pay;
+differentiated pressure must) reproduced on a new modality. Frequency
+reweighting as that pressure
+([exp009](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp009_bandroles))
+was directionally right 4/4 and negligible in magnitude — an honest
+near-collinearity, confessed and rejudged.
+
+The instrument upgrade settled where the payment actually lives.
+[exp010](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp010_controller)
+ships the **StepGatedSampler** — per-step band gating at inference, so
+experts activate coarse-to-fine across the trajectory exactly as
+trained — and judges in image space, where per-step eps-MSE turned out
+to be gauge-blind (0.2% moves there vs these effects). The controller
+lifted round-trip grounding to **+0.2135 vs the frozen trunk's
++0.1247**; band lesions cut grounding monotonically (LOW −0.016 / MID
+−0.031 / HIGH −0.035); and the HIGH-band lesion's image-space
+signature is **14× LP-dominated** (coarse structure) where LOW leans
+detail — the coarse-to-fine thesis confirmed by lesion signatures, not
+by narrative. On real fused data
+([exp011](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp011_fused_multiband),
+2-seed) the whole structural story replicates and the adapters pay
+2–3× more than on synthetic rows.
+
+Then the payer instrument arrived.
+[exp012](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp012_blob_supervised)'s
+in-bed **role-aligned gauge** (HIGH-band foreground-LP error on x0)
+revealed what every aggregate comparison had hidden: **multiband beats
+the matched monolith by ~10% on HIGH-band foreground structure
+(0.1115 vs 0.1234; frozen 0.1555), both seeds.** Band specialists earn
+exactly where the role map says they should — and nowhere an
+aggregate-eps gauge can see.
+
+## The conditioning law
+
+exp012's actual assignment — blob supervision (segmentation-derived
+foreground masks rasterized to the latent grid, coupled to the HIGH
+band) on the eps substrate — returned an honest negative at both
+seeds: +0.03%/−1.0%, inert. The operator's hypothesis said this was
+not a failure of blob supervision but of the *parameterization*:
+epsilon recovers structure through
+`x0 = (x_t − sqrt(1−ᾱ)·ε̂)/sqrt(ᾱ)`, which divides by a vanishing
+`sqrt(ᾱ)` exactly at high noise — exactly where blob supervision
+applies. Flow recovers `x0 = x_t − σ·v` **exactly and linearly at
+every sigma**.
+[exp013](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp013_blob_flow)
+ran the identical coupling on the flow substrate with a preregistered
+5× bar:
+
+> **The conditioning law (2-seed).** Structural supervision pays where
+> x0 recovery is linear and is inert where it is ill-conditioned. The
+> identical blob coupling moved its gauge −5.9%/−3.7% on flow vs
+> +0.03%/−1.0% on eps — a ~125–200× effect ratio, past the
+> preregistered bar at both seeds, with the common gauge inside
+> tolerance.
+
+The three-point dose curve (λ = 0.5 / 1.0 / 2.0 → −5.9%@+0.2% /
+−8.3%@+0.4% / −8.4%@+0.9%) is monotone and saturating: **λ≈1 captures
+the full blob gain inside the common-gauge bound**, and that constant
+now ships as the framework default. A bonus row with its own
+consequence: the undertrained Lune trunk pays adapters −9.1% vs frozen
+— the largest in the line — so adapter headroom scales inversely with
+trunk training. The stated caveat stands: the flow-vs-eps contrast
+rides on different trunks; the v-pred same-trunk arm is the sharpest
+follow-up and is queued, not run.
+
+This is the article's namesake because it is *predictive*: you can
+read the parameterization of a trunk and know, before training,
+whether structure-level supervision has a linear path to pay through.
+In the framework it is encoded literally — `blob=True` with
+`objective="eps"` refuses, with the receipt in the error message.
+
+## Where routing died: three falsifications with teeth
+
+The language line's dispatch rescued specialize-regime anchors. On
+diffusion, comparative routing failed **every** honest test we could
+build, and the failures were as informative as the certifications:
+
+- [exp007](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp007_amoe_lora)
+  (2-seed): a 4-expert bank with dense signed state+sigma dispatch
+  ties/loses to a rank-matched monolith under uniform pressure, with
+  *flat* sigma-band usage (spread 9e-4) — the aggregation channel
+  alone creates no specialization.
+- [exp014](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch)
+  (2-seed): putting text in the dispatch key beats nothing — because
+  the hidden-state key **already routes by prompt** (usage variance
+  2.4e-2 with zero text input); and the raw-flattened [32,128] address
+  *kills* routing outright (1.4e-5) — the language line's
+  high-dimensional disease geometry reproduced on a dispatch key. The
+  bed also confessed its own null was wrong (a shuffled-key null
+  measures diversity, not correctness) and rebuilt it.
+- [exp015](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp015_router_v2)
+  (2-seed, corrected instruments): the M-hat slot-bottlenecked address
+  key — the version the disease-geometry lesson demanded — shows
+  routing excess of **2.5e-06 over a repeated-key null** and a match
+  advantage of ~0. **Address-as-key, in its frozen form, is
+  falsified.** The open form (a text-encoder adapter trained jointly
+  with the dispatch — the operator's "router solidifier") remains
+  queued science.
+
+The address itself is not dead:
+[exp002](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp002_sd15_addrcond)
+showed the frozen address appended to *full text* conditioning is
+redundant (real ≈ deranged) — but the address **alone** steers the
+UNet (+0.029 after 2k steps) and the trunk grew its gates to read it.
+Redundant-in-context, not dead: the redesign target is
+complementarity, injecting where text is not.
+
+The consequence is architectural and encoded: on diffusion,
+`amoe.align()` raises with this record in the error message, and the
+certified alternative to comparative routing is the *structural*
+banding of the previous section. A framework that ships its own
+falsifications as API behavior is the honesty discipline of Parts 2–3
+carried to its conclusion.
+
+## The process is the product
+
+The line kept the house discipline under a four-day clock: every
+major claim 2-seed or explicitly bounded; preregistered bars stated
+before spend (and reported when missed); a label inversion in a
+shipped README corrected and confessed within the hour; an instrument
+gap (reference arms missing role gauges) confessed and rejudged; a
+5.7-hour idle gap confessed in the ledger with the operational law it
+re-earned (never end a turn on an intention — launch first, report
+after); the final replication run disclosed as reserve spend before it
+ran. The complete run record — 47 queue jobs, ~61 GPU-hours at ~97%
+in-window utilization, 27 run logs and the machine-readable ledger —
+is archived with the program's research memory, and every package
+carries its slice.
+
+The productization is part of the record too:
+[amoe-lora 0.2](https://huggingface.co/AbstractPhil/amoe-lora) ships
+the diffusion subsystem with the campaign's laws as constructors (the
+dtype law, the conditioning-law refusal, pure Adam, the toggle and
+zero-init contracts as CPU-testable invariants), the
+[diffusion-pipe fork](https://github.com/AbstractEyes/diffusion-pipe)
+bakes the adapters into its cosmos/anima and SDXL paths (with the
+package's classes when installed, a vendored fallback when not — the
+refactor verified bit-exact against the exp004 checkpoint), and every
+adapter stack in the record now has a bitwise-verified safetensors
+companion, with the curated set in
+[aleph-diffusion-adapters](https://huggingface.co/AbstractPhil/aleph-diffusion-adapters).
+
+## References
+
+External references follow the verified bibliographies of
+[Part 2](https://huggingface.co/AbstractPhil/geolip-aleph-qwen/blob/main/article_ft2.md)
+and [Part 3](https://huggingface.co/AbstractPhil/geolip-aleph-qwen-3.5-0.8b-instruct/blob/main/article_ft3.md);
+the directly load-bearing entries here are:
+
+- Ho, J., Jain, A., & Abbeel, P. (2020). *Denoising Diffusion
+  Probabilistic Models.* [arXiv:2006.11239](https://arxiv.org/abs/2006.11239)
+- Song, J., Meng, C., & Ermon, S. (2020). *Denoising Diffusion
+  Implicit Models.* [arXiv:2010.02502](https://arxiv.org/abs/2010.02502)
+- Rombach, R., et al. (2022). *High-Resolution Image Synthesis with
+  Latent Diffusion Models.* [arXiv:2112.10752](https://arxiv.org/abs/2112.10752)
+- Podell, D., et al. (2023). *SDXL: Improving Latent Diffusion Models
+  for High-Resolution Image Synthesis.* [arXiv:2307.01952](https://arxiv.org/abs/2307.01952)
+- Lipman, Y., et al. (2022). *Flow Matching for Generative Modeling.*
+  [arXiv:2210.02747](https://arxiv.org/abs/2210.02747)
+- Liu, X., Gong, C., & Liu, Q. (2022). *Flow Straight and Fast:
+  Learning to Generate and Transfer Data with Rectified Flow.*
+  [arXiv:2209.03003](https://arxiv.org/abs/2209.03003)
+- Salimans, T., & Ho, J. (2022). *Progressive Distillation for Fast
+  Sampling of Diffusion Models* (the v-parameterization).
+  [arXiv:2202.00512](https://arxiv.org/abs/2202.00512)
+- Hu, E. J., et al. (2021). *LoRA: Low-Rank Adaptation of Large
+  Language Models.* [arXiv:2106.09685](https://arxiv.org/abs/2106.09685)
+- Shazeer, N., et al. (2017). *Outrageously Large Neural Networks: The
+  Sparsely-Gated Mixture-of-Experts Layer.*
+  [arXiv:1701.06538](https://arxiv.org/abs/1701.06538)
+- Puigcerver, J., et al. (2023). *From Sparse to Soft Mixtures of
+  Experts.* [arXiv:2308.00951](https://arxiv.org/abs/2308.00951)
+
+Internal record: the 16 experiment packages under
+[geolip-aleph-diffusion](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion)
+(each with code, `results.json`, and ledger rows; weights in `.pt` +
+verified `.safetensors`), the
+[Part 1](https://huggingface.co/blog/AbstractPhil/aleph-autoregressive-differentiation-ft1),
+[Part 2](https://huggingface.co/AbstractPhil/geolip-aleph-qwen/blob/main/article_ft2.md),
+and [Part 3](https://huggingface.co/AbstractPhil/geolip-aleph-qwen-3.5-0.8b-instruct/blob/main/article_ft3.md)
+articles, the
+[qwen-deepfashion-fused](https://huggingface.co/datasets/AbstractPhil/qwen-deepfashion-fused)
+dataset (the fused-data and blob-target source), and the production
+repos
+([aleph-diffusion-adapters](https://huggingface.co/AbstractPhil/aleph-diffusion-adapters),
+[amoe-lora](https://huggingface.co/AbstractPhil/amoe-lora),
+[diffusion-pipe](https://github.com/AbstractEyes/diffusion-pipe)).