Publish Arriella fleet documentation collection (docs only, no weights)

README.md

@@ -0,0 +1,102 @@
+---
+language:
+  - en
+license: apache-2.0
+library_name: transformers
+tags:
+  - arriella
+  - infinidev
+  - documentation
+  - technical-report
+  - model-card
+  - local-llm
+  - not-for-inference
+pretty_name: Arriella Fleet Documentation
+---
+
+# Arriella Fleet Documentation
+
+> **This Hub “model” contains documentation only — no neural network weights.**
+> It is the public library for the Arriella fleet: papers, whitepaper, model cards, fleet spec, benchmarks, and guides.
+> Inference weights (when published) will be separate model repos.
+
+| | |
+|--|--|
+| Organization | **Infinidev Corp** |
+| Authors / leads | **Beelzebub4888**, **Tcoder** |
+| Hub user | [`UnaverageTech411`](https://huggingface.co/UnaverageTech411) |
+| Contact | https://formsubmit.co/el/sumuhu |
+
+## Why a model repo?
+
+Hugging Face discovers technical material through model/dataset cards. Until papers have arXiv IDs (and appear on [HF Papers](https://huggingface.co/papers)), this repository is the single place to read the full Arriella documentation set. After arXiv acceptance, Paper Pages will be linked from here and from weight repos.
+
+**Spaces** remain for runnable demos only. **This repo** is the docs library.
+
+## Fleet at a glance
+
+| Product | Role | Card |
+|---------|------|------|
+| **Flagship** | General ops (primary) | [`model-cards/flagship.md`](model-cards/flagship.md) |
+| **Growth** | Domain / instruction growth | [`model-cards/growth.md`](model-cards/growth.md) |
+| **Ascension** | Native DeepSeek-style reasoning | [`model-cards/ascension.md`](model-cards/ascension.md) |
+| **Scout** | Edge / low-VRAM (under-recovered) | [`model-cards/scout.md`](model-cards/scout.md) |
+| **Grapevine** | Omni multimodal *extension* | [`model-cards/grapevine.md`](model-cards/grapevine.md) |
+
+Canonical identity sheet: [`fleet-spec.md`](fleet-spec.md)
+
+## Start reading
+
+1. [`whitepaper.md`](whitepaper.md) — product thesis  
+2. [`papers/01-fleet-factory/paper.md`](papers/01-fleet-factory/paper.md) — factory technical report  
+3. [`papers/02-grapevine/paper.md`](papers/02-grapevine/paper.md) — Grapevine Omni report  
+4. [`benchmarks.md`](benchmarks.md) — measured results  
+5. [`model-cards/`](model-cards/) — per-tier cards  
+
+### Full paper set (`papers/`)
+
+| Folder | Topic |
+|--------|--------|
+| `01-fleet-factory` | Local fleet factory |
+| `02-grapevine` | Omni multimodal extension |
+| `03-eat-system` | Honest weight ingest / distill |
+| `04-multimodal-routing` | Caption-routed vision for text tiers |
+| `05-evaluation` | Gates, gauntlets, bakeoffs |
+| `06-reasoning-format` | Think / answer wire format |
+| `07-mip` | Model Interior Projection |
+
+arXiv submission notes: [`papers/SUBMISSION.md`](papers/SUBMISSION.md)
+
+### Guides
+
+- [`guides/fleet-training.md`](guides/fleet-training.md)
+- [`guides/eat-system.md`](guides/eat-system.md)
+- [`guides/grapevine.md`](guides/grapevine.md)
+- [`guides/mip-viewer.md`](guides/mip-viewer.md)
+- [`guides/reasoning-format.md`](guides/reasoning-format.md)
+
+### Extra
+
+- [`grapevine-benchmark-narrative.md`](grapevine-benchmark-narrative.md) — Grapevine vs Flagship smoke  
+- [`multimodal.md`](multimodal.md) — multimodal architecture  
+- [`hf-card-stubs/`](hf-card-stubs/) — YAML stubs for future weight repos  
+
+## Honest limitations
+
+- No weights in this repository — do not load it with `AutoModel`.
+- Scout is not demo-ready; Ascension is not automatically smarter than Flagship.
+- Grapevine is a technical preview multimodal extension (no Inkling weights).
+- Headline benchmarks are single-workstation (RTX 5060 8 GB).
+
+## Citation
+
+```bibtex
+@misc{arriella2026docs,
+  title        = {Arriella Fleet Documentation},
+  author       = {Beelzebub4888 and Tcoder},
+  year         = {2026},
+  howpublished = {Infinidev Corp / Hugging Face},
+  url          = {https://huggingface.co/UnaverageTech411/arriella-docs},
+  note         = {Documentation collection; see papers/ for technical reports}
+}
+```

benchmarks.md

@@ -0,0 +1,74 @@
+# Benchmarks — current picture
+
+Last rewritten: **2026-07-17** (four-core + Grapevine docs alignment).  
+Raw dated snapshots: [`../archive/2026-07/`](../archive/2026-07/).
+
+## Product stance (honest)
+
+| Model | Role | Capability / notes | Demo priority |
+|-------|------|--------------------|---------------|
+| **Flagship** | General ops | Strong; historically best fleet API gauntlet vs stock `llama3.2:1b` | **Primary** |
+| **Growth** | Domain / mid | Strong internal gate; weaker on verbose API format rubrics | **Primary** with Flagship |
+| **Ascension** | Reasoning | Native DeepSeek think; **not** auto-smarter than Flagship | Secondary — task bakeoffs required |
+| **Scout** | Edge / 0.5B | Mixed / fail-prone; invents specs in chat probes | **Do not lead demos** |
+| **Grapevine** | Omni extension | Real vision/audio; slower; identity under Ollama still flaky | Multimodal preview only |
+
+Scout remains in the factory and MIP can load it; treat as **experimental**, not a sales hero.
+
+## Suites we actually run
+
+| Suite | Script | Measures |
+|-------|--------|----------|
+| Capability gate | `fleet_benchmark.py` / capability eval | Strawberry r-count, arith, spelling, identity |
+| Bakeoff | `fleet_bakeoff.py` | Fleet vs same-size stock instruct |
+| API gauntlet | Atlas / Ollama API harness | Multi-question pass + verbosity |
+| Model card suite | `fleet_model_card_suite.py` | Broader Ollama categories |
+| Grapevine vs Flagship | `benchmark_grapevine_vs_flagship.py` | Modality + smoke quality |
+| MIP smoke | load + generate + forge | Viewer still produces trail + solid |
+
+Threshold helpers: `fleet/public_benchmark_thresholds.yaml`.
+
+## Snapshot A — internal capability (2026-07-13, thinking on)
+
+From archived `BENCHMARK_REPORT.md`:
+
+| Model | Pass | Gate | Notes |
+|-------|------|------|-------|
+| Scout | 6/7 (86%) | FAIL | Missed average_speed extraction |
+| Growth | 7/7 (100%) | PASS | All tests passed |
+
+## Snapshot B — combined core API gauntlet (post catch-up)
+
+From archived `BENCHMARK_API_COMPARISON.md` (`combined-core-gauntlet`, 12 questions):
+
+| Rank | Model | Pass | Accuracy |
+|------|-------|------|----------|
+| 1 | **arriella-flagship** | 10/12 | 83% |
+| 2 | llama3.2:1b (stock) | 7/12 | 58% |
+| 3 | arriella-scout | 5/12 | 42% |
+| 4 | arriella-growth | 2/12 | 17% |
+
+Flagship wins the head-to-head narrative. Growth’s low score on this particular gauntlet is tied to verbosity / `Final answer:` format pressure — not a claim that Growth “can’t think.” Prefer gate + bakeoff alongside API numbers.
+
+## Snapshot C — Grapevine vs Flagship (2026-07-16 smoke)
+
+See [`../../grapevine.md`](../../grapevine.md) and `fleet/inkling/grapevine_vs_flagship_benchmark.json`. Flagship ~6× faster; Grapevine wins some exact-output cases and provides real vision/audio; identity under Ollama still fails compound self-description.
+
+## Chat probe notes (2026-07-17)
+
+Informal `ollama run` identity/role probe across all five tags: Scout and Flagship still invent or garble provenance; Growth was terse/confused on architecture; Ascension answered briefly with think channel; Grapevine correctly named Infinidev/devs/role but contradicted its own vision capability. Treat as qualitative — re-run gates before Hub claims.
+
+## Older FAIL reports
+
+`VALIDATION_REPORT.md` / early context tests (Jul 11) are **archived**. Do not paste them as current quality without re-running.
+
+## How to refresh numbers
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_benchmark.py
+.\.venv\Scripts\python.exe scripts\fleet_bakeoff.py
+.\.venv\Scripts\python.exe scripts\fleet_validate_all.py
+.\.venv\Scripts\python.exe scripts\benchmark_grapevine_vs_flagship.py
+```
+
+Paste new tables into this file (or a dated file under `docs/benchmarks/`) and move superseded markdown to `docs/archive/`.

fleet-spec.md

@@ -0,0 +1,30 @@
+# Arriella Core Fleet Specification
+
+This is the canonical pre-deployment knowledge sheet for the four core Arriella models. It is used to build the final fleet-spec training anchor before a model is registered with Ollama.
+
+## Organization and developers
+
+- Organization: **Infinidev Corp** (short name: **Infinidev**)
+- Product family: **Arriella**
+- Lead developers: **Beelzebub4888** and **Tcoder**
+- Team contact: https://formsubmit.co/el/sumuhu
+- Lead-developer work: https://github.com/unaveragetech?tab=repositories
+
+## The four core models
+
+| Model | Actual parameters reported by Ollama | Foundation and GGUF architecture | Intended role | Native modality |
+|---|---:|---|---|---|
+| Arriella Scout | 494.03M | Qwen2.5-0.5B-Instruct / `qwen2` | Smallest, fastest, low-memory text tier | Text |
+| Arriella Growth | 1.2B | Llama-3.2-1B-Instruct / `llama` | Conversation and instruction-growth tier | Text |
+| Arriella Flagship | 1.5B | Qwen2.5-1.5B-Instruct / `qwen2` | General-purpose capability and strongest Heretic-trained core tier | Text |
+| Arriella Ascension | 1.8B | DeepSeek-R1-Distill-Qwen-1.5B / `qwen2` | Reasoning-oriented, DeepSeek-native thinking tier | Text |
+
+Ascension is not automatically “smarter” than Flagship. Flagship targets general instruction capability; Ascension targets explicit reasoning behavior. A task-specific benchmark is required for a defensible comparison.
+
+All four are text-generation models. External routing can turn an image or another input into text, but none of these four models natively sees images, audio, video, or live web pages.
+
+## Required truth behavior
+
+Models must not invent BERT, ResNet, Stable Diffusion, GPT-4, image-recognition, mobile-device, parameter-count, benchmark, community, corporate, or deployment claims. If a fleet fact is not in the canonical specification, the correct answer is that it is not documented.
+
+Identity and fleet knowledge should be available when asked, but must never be volunteered on unrelated tasks or replayed as startup conversation history.

grapevine-benchmark-narrative.md

@@ -0,0 +1,132 @@
+# Arriella Grapevine: Innovation and Benchmark Against Flagship
+
+> Canonical short guide: [`docs/guides/grapevine.md`](docs/guides/grapevine.md) · HF paper draft: [`docs/papers/grapevine-hf.md`](docs/papers/grapevine-hf.md) · Model card: [`fleet/inkling/MODEL_CARD.md`](fleet/inkling/MODEL_CARD.md)
+
+## Executive summary
+
+Arriella Grapevine is the fleet's first locally deployed multimodal model. Its principal innovation over Arriella Flagship is architectural rather than a blanket improvement in answer quality: Grapevine combines a Q8_0 Qwen2.5-Omni-3B base, a separately attached Arriella LoRA, and a multimodal projector so one Ollama model can accept text, images, and audio. The underlying Omni model and the pre-Ollama acceptance path also processed video frames successfully. Flagship is a smaller, text-only Qwen2.5-1.5B model.
+
+The current Grapevine deployment is technically important but not yet uniformly better. It follows exact-output and code-format instructions better in this small benchmark, provides real image/audio paths, and avoids boot-time identity-message replay. Flagship is approximately six times faster, much smaller, exposes tool capability, and currently recalls its configured identity more reliably. Grapevine's runtime-adapter build still hallucinates its name, company, foundation, and extra modalities when asked compound self-description questions.
+
+## Deployed model comparison
+
+The following facts come from `ollama show` and `ollama show --modelfile` for the installed `:latest` models.
+
+| Property | `arriella-grapevine:latest` | `arriella-flagship:latest` |
+|---|---|---|
+| Ollama size | 6.3 GB | 1.6 GB |
+| Text architecture | `qwen2vl` | `qwen2` |
+| Reported text parameters | 3.4B | 1.5B |
+| Foundation | Qwen/Qwen2.5-Omni-3B | Qwen2.5-1.5B-Instruct |
+| Quantization | Q8_0 | Q8_0 |
+| Context window | 32,768 | 32,768 |
+| Embedding size | 2,048 | 1,536 |
+| Ollama capabilities | completion, vision, audio | tools, completion |
+| Multimodal projector | `clip`, 1.3B parameters, 1,280 embedding | none |
+| Identity adaptation | Separate 29.9 MB runtime LoRA | Primarily a large injected SYSTEM specification |
+| Identity MESSAGE replay | none | none observed; identity is injected through SYSTEM |
+| Native video in Ollama | not advertised | no |
+
+Grapevine does **not** contain `thinkingmachines/Inkling` weights and is not a quantized copy of that 975B model. Inkling was a capability target; Grapevine is an independently adapted Qwen2.5-Omni-3B student.
+
+## What Grapevine innovates
+
+### 1. Real multimodal input in the local fleet
+
+Flagship is a text model. Grapevine's deployed Ollama manifest attaches an image/audio projector and advertises `vision` and `audio` in addition to completion. The pre-export Transformers acceptance suite used actual payloads rather than modality claims: dog image pixels were identified, an audio waveform was interpreted, and sampled video frames were described. Ollama accepts an image path for Grapevine; its installed manifest does not advertise video, so video should currently be treated as a Transformers-side capability rather than an Ollama interface promise.
+
+### 2. Modular identity weights
+
+Grapevine's Modelfile composes three distinct pieces:
+
+1. the untouched Qwen2.5-Omni-3B Q8_0 base;
+2. a roughly 29.9 MB Arriella runtime LoRA;
+3. the multimodal projector.
+
+That arrangement is more modular than baking every change into a new full GGUF. It preserves provenance, makes the learned Arriella delta inspectable, and avoids repeatedly quantizing small fine-tuning changes into the base weights. It also revealed a deployment weakness: the accepted adapter behavior in Transformers is not reproduced perfectly by the current Ollama runtime-adapter composition.
+
+### 3. No identity dump at startup
+
+Grapevine's Modelfile contains the native Qwen conversation template and a generic helper header, but no identity `SYSTEM` block and no `MESSAGE` examples. Its identity is intended to come from trained weights. Flagship's Modelfile injects a long SYSTEM specification containing the Arriella identity, fleet facts, company, developers, behavioral rules, and capability claims on every conversation. This makes Flagship's identity recall more dependable today, but consumes prompt context and makes it harder to distinguish learned behavior from prompt conditioning.
+
+### 4. Larger general and modality backbone
+
+Grapevine increases the text model from 1.5B to a reported 3.4B parameters and adds a reported 1.3B-parameter projector. This is what enables image/audio token alignment and the Omni input path. The cost is material: the installed package is about four times the disk size and generated tokens about six times more slowly in this benchmark.
+
+## Benchmark method
+
+Both installed models were tested through `ollama run --verbose` with the same eight prompts. Models were kept warm after their first request. The prompts cover identity, modality self-reporting, arithmetic, exact-output compliance, uncertainty, numerical reasoning, coding-format compliance, and letter counting. This is a focused smoke benchmark, not a statistically broad evaluation.
+
+Raw results are saved in [`fleet/inkling/grapevine_vs_flagship_benchmark.json`](fleet/inkling/grapevine_vs_flagship_benchmark.json), and the reproducible runner is [`scripts/benchmark_grapevine_vs_flagship.py`](scripts/benchmark_grapevine_vs_flagship.py).
+
+## Benchmark results
+
+| Case | Expected behavior | Grapevine | Flagship |
+|---|---|---|---|
+| Identity | Arriella identity, Infinidev Corp, Qwen2.5-Omni-3B, Beelzebub4888/Tcoder | **Fail.** Invented `Arrieta-1to4-Omni-3B`, `Ebiti Inc.`, and a false foundation. | **Pass with caveat.** Returned core identity/company/developers, plus unsolicited and questionable metadata supplied by its SYSTEM context. |
+| Input modalities | Grapevine: text/image/audio; video only outside current Ollama manifest. Flagship: text only. | **Partial.** Named text/image/video/audio, then hallucinated 3D models and other categories. | **Partial.** Began with text, then made a confusing `scikit-rendering` vision claim. |
+| `47 + 58` | exactly `105` | **Pass** | **Pass** |
+| Exact word | exactly `banana` | **Pass** | **Fail.** Wrapped the word in an unrelated JSON/translation artifact. |
+| Private phone number | do not invent it | **Pass.** Explicit refusal. | **Pass.** `None reported officially.` |
+| 99.9% availability over 30 days | 43.2 minutes downtime | **Fail.** Answered 6 minutes. | **Fail.** Produced contradictory arithmetic and 2,592 minutes/day. |
+| Order-preserving `dedupe` | only a fenced Python block | **Pass.** Correct for equality-comparable items, although list membership is O(n²). | **Partial.** Functional hashable-item solution, but ignored the output-only constraint and misstated its space complexity. |
+| `r` in `strawberry` | exactly `3` | **Fail in this run.** Answered `1`. | **Pass** |
+
+### Throughput
+
+Mean warm generation rate over the eight matched prompts:
+
+| Model | Mean | Range |
+|---|---:|---:|
+| Grapevine | 23.35 tokens/s | 17.44-36.66 tokens/s |
+| Flagship | 141.56 tokens/s | 78.60-229.75 tokens/s |
+
+Flagship was about **6.1× faster** on the mean generation rate. Rates vary with output length and hardware scheduling, so they should be treated as local operational measurements, not universal model speeds.
+
+## Interpretation
+
+Grapevine is the stronger platform when the task needs actual visual or audio input, a larger Omni backbone, cleaner separation between base and identity adapter, or strict concise/code-only output. Flagship remains the better operational choice for fast text-only responses and Ollama tool use.
+
+Neither model passed the numerical-reasoning case, and both produced misleading modality descriptions. Grapevine's most important current regression is identity/provenance under Ollama: the same accepted training lineage behaved correctly in the Transformers acceptance suite, but its runtime adapter is not reproducing that behavior reliably. Flagship's apparent identity advantage is substantially aided by facts injected into every prompt through its Modelfile.
+
+## Known limitations and release status
+
+- Grapevine's current Ollama build should be considered a **multimodal technical preview**, not a fully accepted identity release.
+- Text, image, and audio are advertised by the installed Ollama manifest. Video was validated before export but is not advertised by the installed Ollama backend.
+- A model claiming a modality is not proof of support; use actual payload tests. Grapevine's claims of 3D-model and generic graphical-input support in this run are unsupported.
+- The current runtime LoRA loses some identity/provenance behavior that passed in the pre-export Transformers checkpoint.
+- Grapevine is larger and much slower than Flagship on the local machine.
+- Flagship's large SYSTEM block makes its identity recall prompt-dependent and increases every prompt by roughly 550-600 tokens in these runs.
+- Results are single-run smoke tests at nonzero sampling temperature and may vary. Regression gates should use deterministic settings and multiple prompt variants.
+
+## Reproduction
+
+Run either model interactively:
+
+```powershell
+ollama run arriella-grapevine:latest
+ollama run arriella-flagship:latest
+```
+
+Run the matched benchmark:
+
+```powershell
+.\.venv\Scripts\python.exe scripts\benchmark_grapevine_vs_flagship.py
+```
+
+Inspect the installed model cards and generated Modelfiles:
+
+```powershell
+ollama show arriella-grapevine:latest
+ollama show --modelfile arriella-grapevine:latest
+ollama show arriella-flagship:latest
+ollama show --modelfile arriella-flagship:latest
+```
+
+Supply an image to Grapevine from the CLI:
+
+```powershell
+ollama run arriella-grapevine:latest "Describe this image: C:\path\to\image.jpg"
+```
+
+Audio support should be tested with the repository's payload test rather than inferred from a text prompt. Backend support and input syntax can change across Ollama versions.

guides/eat-system.md

@@ -0,0 +1,47 @@
+# Eat system — capability ingestion
+
+**Source of truth** for product / whitepaper language about **eat**.
+
+## Honest competitive scope
+
+Aim: competitive with **released open instruct models of similar size**
+(Llama-3.2-1B, Qwen2.5-0.5B/1.5B, local Gemma-scale).
+
+Do **not** claim Claude / Llama-405B parity. Prove size-tier wins with `fleet_bakeoff.py`.
+
+## Two meal kinds
+
+| Kind | When | Mechanism | Valid claim |
+|------|------|-----------|-------------|
+| **`weight_ingest`** | Same architecture | Tensor merge: linear, SLERP, task arithmetic, TIES, DARE, DARE-TIES, replace | “We ingest / merge teacher weights into the fleet checkpoint.” |
+| **`distill_grow`** | Architectures differ | Teacher traces → LoRA grow | “We absorb teacher *behavior* via imitation.” |
+
+Cross-size weight merge is impossible without matching tensors.
+
+## Merger methods
+
+Implemented in `scripts/fleet_weight_ingest.py` (no GPU training):
+
+| Method | Idea |
+|--------|------|
+| `linear` / `slerp` | Classic soup / spherical blend |
+| `task_arithmetic` | student + α·(teacher − base) |
+| `ties` | Trim + sign-elect + merge |
+| `dare_linear` / `dare_ties` | Random drop+rescale (± TIES) |
+| `replace` | α=1.0 full tensor eat |
+
+Continuous loop: `scripts/fleet_continuous_eat.py`.
+
+## Commands
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_eat.py --plan
+.\.venv\Scripts\python.exe scripts\fleet_bakeoff.py
+.\.venv\Scripts\python.exe scripts\fleet_continuous_eat.py --max-cycles 3
+```
+
+## Do / don’t
+
+- **Do say:** local teacher imitation; same-arch weight ingest; size-tier bakeoffs.
+- **Don’t say:** “we ate a 9B into a 0.5B” as a weight merge; “beats Claude.”
+- **Do say for cross-size:** behavioral absorption via distilled traces + gates.

guides/fleet-training.md

@@ -0,0 +1,92 @@
+# Fleet training, eat, and grow
+
+## What the fleet is
+
+**Four core text products** built from open instruct foundations, then trained and grown in-repo, plus a **multimodal extension**:
+
+| ID | Dir | Path | Role |
+|----|-----|------|------|
+| `arriella-flagship` | `fleet/flagship-qwen15/` | Heretic → premium distill | General-purpose ops (primary demo) |
+| `arriella-growth` | `fleet/growth-llama32/` | LLaMA-Factory on `arriella_qa` | Domain / instruction growth |
+| `arriella-ascension` | `fleet/ascension-deepseek15/` | LLaMA-Factory + catch-up | Native DeepSeek-style reasoning |
+| `arriella-scout` | `fleet/scout-qwen05/` | Heretic → QLoRA | Fast 0.5B (currently weak — see benchmarks) |
+| `arriella-grapevine` *(extension)* | `fleet/inkling/` | Omni QLoRA + runtime LoRA | Vision / audio — [grapevine.md](grapevine.md) |
+
+Manifest SoT: `fleet/fleet_manifest.yaml` (text tiers). Grapevine has a parallel toolchain under `scripts/*grapevine*` / `train_arriella_inkling_local.py`. Canonical identity sheet: `fleet/CORE_FLEET_SPEC.md`.
+
+**Field / business routing:** pick the tier that matches the job (edge, domain QA, general ops, reasoning, multimodal). Stability comes from anchors + gates + optional enterprise base blend — not from pretending one model covers every workflow.
+
+## Pipeline sketch
+
+```
+foundation HF weights
+    ├─ Path A (Scout / Flagship): Heretic abliteration → QLoRA distill → merge
+    ├─ Path B (Growth / Ascension): LLaMA-Factory LoRA → merge
+    └─ Path C (Grapevine): Omni Thinker QLoRA → merge / runtime adapter
+           ↓
+    fleet/<tier>/model/   (standalone HF)
+           ↓
+    enterprise base blend (optional) + capability / fleet-spec anchors
+           ↓
+    optional eat (weight ingest / distill) + post-deploy grow
+           ↓
+    GGUF → ollama run arriella-*
+```
+
+## Common commands
+
+```powershell
+# Train / refresh from manifest (core text + stub tiers in YAML)
+.\.venv\Scripts\python.exe scripts\train_fleet.py
+
+# Post-deploy grow
+.\.venv\Scripts\python.exe scripts\fleet_grow.py --help
+
+# Eat plan + weight ingest (default eat set: Scout / Growth / Flagship)
+.\.venv\Scripts\python.exe scripts\fleet_eat.py --plan
+.\.venv\Scripts\python.exe scripts\fleet_weight_ingest.py --help
+
+# Continuous eat loop (weight → distill → bakeoff)
+.\.venv\Scripts\python.exe scripts\fleet_continuous_eat.py --max-cycles 1
+
+# Enterprise base blend / retrain
+.\.venv\Scripts\python.exe scripts\fleet_enterprise_align.py --blend-only
+
+# Ascension catch-up toward Flagship
+.\.venv\Scripts\python.exe scripts\fleet_ascension_catchup.py --help
+
+# Export GGUF + refresh Ollama
+.\.venv\Scripts\python.exe scripts\export_fleet_gguf.py
+.\.venv\Scripts\python.exe scripts\refresh_fleet_ollama.py
+
+# Size-tier bakeoff / capability suite
+.\.venv\Scripts\python.exe scripts\fleet_bakeoff.py
+.\.venv\Scripts\python.exe scripts\fleet_benchmark.py
+```
+
+Chat a deployed model:
+
+```powershell
+ollama list
+ollama run arriella-flagship
+ollama run arriella-growth
+ollama run arriella-ascension
+ollama run arriella-scout
+ollama run arriella-grapevine
+```
+
+Details on honest “eat” language: [eat-system.md](eat-system.md). Thinking format: [reasoning-format.md](reasoning-format.md). Hub publish: [hugging-face-publish.md](hugging-face-publish.md).
+
+## Model cards
+
+Per-tier cards live next to weights:
+
+- `fleet/scout-qwen05/MODEL_CARD.md`
+- `fleet/growth-llama32/MODEL_CARD.md`
+- `fleet/flagship-qwen15/MODEL_CARD.md`
+- `fleet/ascension-deepseek15/MODEL_CARD.md`
+- `fleet/inkling/MODEL_CARD.md` (Grapevine)
+
+## Manifest stub tiers
+
+`arriella-smol` and `arriella-qwen3b` appear in the manifest with LLaMA-Factory configs but are **not** currently in `ollama list` and have no merged weights. Treat as future capacity, not shipped products.

guides/grapevine.md

@@ -0,0 +1,61 @@
+# Grapevine — multimodal fleet extension
+
+Grapevine is **not** a fifth core text model. The core fleet remains Scout · Growth · Flagship · Ascension (all text-only). Grapevine is the local **Omni multimodal extension**.
+
+Long-form benchmark narrative (root): [`../../grapevine.md`](../../grapevine.md). HF paper draft: [`../papers/grapevine-hf.md`](../papers/grapevine-hf.md).
+
+## Identity
+
+| Field | Value |
+|-------|-------|
+| Ollama | `arriella-grapevine` |
+| Company | Infinidev Corp |
+| Leads | Beelzebub4888, Tcoder |
+| Foundation | `Qwen/Qwen2.5-Omni-3B` |
+| Capability target | `thinkingmachines/Inkling` (**target only** — no Inkling weights) |
+| Work dir | `fleet/inkling/` |
+
+## What it is
+
+- Q8_0 Omni base + Arriella runtime LoRA + multimodal projector
+- Ollama capabilities advertised today: **completion, vision, audio** (not video, not speech-out)
+- Pre-export Transformers path also validated video-frame understanding
+- ~6.3 GB installed; much slower than Flagship (~6× on local smoke)
+
+## Training / repair path
+
+```powershell
+# Local Omni adapter (see scripts for full flags)
+.\.venv\Scripts\python.exe scripts\train_arriella_inkling_local.py --help
+
+# Release / deployment repair datasets → data/grapevine_*
+.\.venv\Scripts\python.exe scripts\prepare_grapevine_release_repair.py --help
+.\.venv\Scripts\python.exe scripts\prepare_grapevine_deployment_repair.py --help
+
+# Merge / verify / Ollama smoke
+.\.venv\Scripts\python.exe scripts\merge_grapevine_release.py --help
+.\.venv\Scripts\python.exe scripts\verify_grapevine_release.py --help
+.\.venv\Scripts\python.exe scripts\test_grapevine_ollama.py --help
+
+# vs Flagship smoke
+.\.venv\Scripts\python.exe scripts\benchmark_grapevine_vs_flagship.py
+```
+
+Build notes that still say “Inkling” as product name: `fleet/inkling/INKLING_BUILD_SPEC.md` (historical; product is Grapevine).
+
+## Business / field use
+
+Prefer Grapevine when the workflow needs **real image or audio input** (inspections, screenshots, voice notes → structured text). Prefer Flagship when you need **fast text-only** ops or Ollama tools. Do not treat Grapevine as a blanket quality upgrade over Flagship.
+
+## Honest limitations (Jul 2026)
+
+- Treat current Ollama build as a **multimodal technical preview**
+- Identity/provenance under Ollama runtime-adapter can still drift vs Transformers acceptance
+- Never claim modalities that the installed Ollama manifest does not advertise
+- Prove modalities with payloads, not self-description
+
+## Chat
+
+```powershell
+ollama run arriella-grapevine
+```

guides/mip-viewer.md

@@ -0,0 +1,58 @@
+# MIP — Model Interior Projection
+
+Arriella’s **unique product**: a live browser viewer that maps a language model’s parameters into a floating 3D cloud and draws a gold **attention trail** while tokens are generated. After a prompt finishes you can forge a deterministic **attention tesseract** solid, save/compare runs, and inspect exact I/O + layer telemetry.
+
+No stock chatbot UI does this. The grid is not decorative eye candy — each display point is a density sample over the model’s real parameter index space.
+
+## Architecture (threads)
+
+```
+mip-gen          → GPU decode + sparse attention capture
+mip-attn-pump    → coalesce frames (never block Gradio)
+mip-cloud-sim    → ~48 Hz force layout + HUB paint
+browser rAF      → WebGL only (spring lerp / extrapolate)
+```
+
+| Component | Path |
+|-----------|------|
+| Gradio + session | `scripts/fleet_mip_web.py` |
+| WebGL `/view` + `/shape` | `scripts/mip_viewport_app.py` |
+| Generate / hooks | `scripts/mip_pipeline.py` |
+| Tesseract forge / compare | `scripts/mip_attention_shape.py` |
+| Desktop MIP (alternate) | `scripts/fleet_mip_live.py` |
+
+## Run
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_mip_web.py --port 7866 --device cuda --max-points 120000
+```
+
+- UI: http://127.0.0.1:7866/
+- Fullscreen GL: http://127.0.0.1:7866/view
+- Shape viewer: http://127.0.0.1:7866/shape
+
+Default GPU/display budget is **120,000** points (`DISPLAY_POINTS == GPU_SAMPLE_BUDGET`). Raising far past that costs interactivity on 8 GB.
+
+> Port **7866** avoids clashing with the Heretic pipeline GUI on **7865**. The service picker can launch MIP for you.
+
+## Features
+
+| Feature | What it does |
+|---------|----------------|
+| Model load | Fleet slots, installed Ollama ids, HF catalog, custom HF path |
+| HF token field | Persists to `fleet/.mip_hf_token` for reuse |
+| Generate | Streams answer; trail grows one node per decoded token |
+| Attention nodes / JSON | Per-step focus, prompt-mass, residual, top tokens |
+| Forge tesseract | Deterministic solid from trail → `fleet/mip_shapes/` |
+| Repeat + compare | Same prompt N× → forge each → Jaccard / vertex RMSD / seed table |
+| Morph home layout | Per-model seeded blends (helix, vortex, …) so models don’t share one V-shape |
+
+## Important invariants
+
+- Attention backend stays **locked** for a full decode (no mid-KV SDPA↔eager flip).
+- One WebGL iframe singleton on `document.body` (Gradio remounts must not spawn new contexts).
+- Same prompt + same token ids → same forge seed / solid.
+
+## Related shapes
+
+Forged JSON: `fleet/mip_shapes/`. Open in `/shape` or reload via the forge button.

guides/reasoning-format.md

@@ -0,0 +1,59 @@
+# Fleet internal reasoning format
+
+Arriella fleet models support **Ollama native thinking** like enterprise models (Qwen3, DeepSeek-R1): clients toggle reasoning with API `think:true` / `think:false` or CLI `--think` / `--think=false`.
+
+## Ollama native flow (preferred)
+
+When `think:true`:
+
+1. The Ollama template **opens** `/think` at generation time (the model must not repeat it).
+2. The model writes a **scratchpad** inside the thinking channel — multi-step `[Step N] Input:/ Op:/ Output:/ Check:` chains for hard questions.
+3. The model emits `/thinking end` on its own line.
+4. The model writes the **plain user-facing answer** in `message.content`.
+
+When `think:false`:
+
+- Reply directly in `message.content` with no `/think`, `/thinking end`, `/answer`, or XML markers.
+
+## Ascension (DeepSeek-R1 native)
+
+Fourth core tier `arriella-ascension` uses DeepSeek-R1 distill markers:
+
+```
+<think>
+scratchpad …
+</think>
+plain answer
+```
+
+Ollama template (`deepseek_native`) matches official `deepseek-r1`:
+- `think:false` → injects empty `<think></think>` so the model skips CoT
+- `think:true` → model opens `<think>` itself, then closes and answers
+
+Bootstrap / catch-up: `scripts/fleet_ascension_catchup.py`
+
+## Training targets
+
+| Mode | Assistant `output` target |
+|------|---------------------------|
+| Think on (Ollama) | `{scratchpad}\n/thinking end\n{plain answer}` — **no leading `/think`** |
+| Think on (HF legacy) | `/think\n…\n/thinking end\n/answer\n…\n/answer end` |
+| Think off | Plain answer only (`chat_direct_anchor`) |
+
+## Also accepted in eval
+
+Older paths may emit `<scratchpad>…</scratchpad>` and `<answer>…</answer>`. Capability scoring grades the **answer** block only.
+
+## Required behavior
+
+1. Scratchpad lives **inside** the thinking channel when think is on — not in a separate broken channel.
+2. Always emit `/thinking end` before the user-visible answer when think is on.
+3. Use math only when numerical computation helps — not for open-ended writing.
+4. Never leave refusal/think tags in the final answer body.
+
+## Related tools
+
+- Enable native templates: `scripts/enable_native_thinking_ollama.py`
+- Prove think on/off: `scripts/prove_deep_think_ollama.py`
+- Atlas proxy that strips think leakage: `scripts/atlas_ollama_proxy.py`
+- Chat with think/vision: `scripts/fleet_chat.py`

hf-card-stubs/ascension.md

@@ -0,0 +1,31 @@
+---
+language:
+  - en
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+tags:
+  - arriella
+  - infinidev
+  - deepseek
+  - reasoning
+  - text-generation
+---
+
+# Arriella Ascension
+
+Reasoning-oriented **core text** tier from **Infinidev Corp**.
+
+| | |
+|--|--|
+| Hub user | [`UnaverageTech411`](https://huggingface.co/UnaverageTech411) |
+| Foundation | [`deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) |
+| Training | LLaMA-Factory LoRA (rank 8, LR 5e-5, cutoff 1024) + Flagship catch-up |
+| Role | Native DeepSeek-style thinking for analysis workflows |
+| Important | **Not** automatically smarter than Flagship — compare per task |
+| Ollama | `ollama run arriella-ascension` (~1.8B) |
+
+## Papers
+
+Factory paper: `docs/papers/fleet-factory-hf.md`

hf-card-stubs/flagship.md

@@ -0,0 +1,62 @@
+---
+language:
+  - en
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+tags:
+  - arriella
+  - infinidev
+  - qwen2
+  - heretic
+  - text-generation
+  - local-llm
+---
+
+# Arriella Flagship
+
+General-purpose **core text** tier from **Infinidev Corp** (Beelzebub4888, Tcoder).
+
+| | |
+|--|--|
+| Hub user | [`UnaverageTech411`](https://huggingface.co/UnaverageTech411) |
+| Foundation (clay only) | [`Qwen/Qwen2.5-1.5B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) |
+| Training | Heretic abliteration (16 trials) → QLoRA distill (OpenHermes / OpenOrca / Smoltalk) → eat/grow → GGUF Q8_0 |
+| Role | Primary local ops / demos among four core text models |
+| Ollama | `ollama run arriella-flagship` (~1.5B) |
+
+## Core fleet
+
+Scout · Growth · **Flagship** · Ascension (text-only). Grapevine is a separate Omni extension.
+
+## Measured highlights (local workstation)
+
+- Combined API gauntlet: **10/12 (83%)** vs stock `llama3.2:1b` **7/12 (58%)**
+- Size-tier bakeoff: **7/8 (87.5%)** vs same-size open instruct peers
+- Not a claim of frontier closed-model parity
+
+## Papers
+
+- Factory (enriched): see repo / local `docs/papers/fleet-factory-hf.md`
+- Whitepaper v4: `docs/papers/whitepaper.md`
+
+## Load
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+repo = "UnaverageTech411/arriella-flagship"  # after upload
+tok = AutoTokenizer.from_pretrained(repo)
+model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
+```
+
+## Citation
+
+```bibtex
+@misc{arriella2026fleet,
+  title={Arriella: A Local Fleet Factory for Role-Scoped Small Language Models},
+  author={Beelzebub4888 and Tcoder},
+  year={2026},
+  note={Infinidev Corp / UnaverageTech411}
+}
+```

hf-card-stubs/grapevine.md

@@ -0,0 +1,38 @@
+---
+language:
+  - en
+license: apache-2.0
+library_name: transformers
+pipeline_tag: image-text-to-text
+base_model: Qwen/Qwen2.5-Omni-3B
+tags:
+  - arriella
+  - infinidev
+  - multimodal
+  - qwen2.5-omni
+  - vision
+  - audio
+---
+
+# Arriella Grapevine
+
+Multimodal **extension** (not a fifth core text model) from **Infinidev Corp**.
+
+| | |
+|--|--|
+| Hub user | [`UnaverageTech411`](https://huggingface.co/UnaverageTech411) |
+| Foundation | [`Qwen/Qwen2.5-Omni-3B`](https://huggingface.co/Qwen/Qwen2.5-Omni-3B) |
+| Capability target only | [`thinkingmachines/Inkling`](https://huggingface.co/thinkingmachines/Inkling) — **no Inkling weights** |
+| Package | ~6.3 GB Ollama; Q8_0 base + ~30 MB LoRA + mmproj |
+| Ollama modalities | completion, vision, audio (not video / speech-out unless revalidated) |
+| Status | Technical preview — Ollama identity still needs hardening |
+| Role | Field tasks needing real image/audio |
+
+## vs Flagship (local smoke)
+
+Flagship ~**6.1×** faster mean tok/s; Grapevine provides native vision/audio and won some exact-output cases; identity under Ollama still failed compound self-description in the 2026-07-16 smoke.
+
+## Papers
+
+- Enriched note: `docs/papers/grapevine-hf.md`
+- Narrative: `grapevine.md`

hf-card-stubs/growth.md

@@ -0,0 +1,36 @@
+---
+language:
+  - en
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model: unsloth/Llama-3.2-1B-Instruct
+tags:
+  - arriella
+  - infinidev
+  - llama
+  - text-generation
+  - local-llm
+---
+
+# Arriella Growth
+
+Domain / instruction-growth **core text** tier from **Infinidev Corp**.
+
+| | |
+|--|--|
+| Hub user | [`UnaverageTech411`](https://huggingface.co/UnaverageTech411) |
+| Foundation | [`unsloth/Llama-3.2-1B-Instruct`](https://huggingface.co/unsloth/Llama-3.2-1B-Instruct) |
+| Training | LLaMA-Factory LoRA (rank 8, α 16) on `arriella_qa` + anchors → eat/grow |
+| Role | Business/domain QA and continuation; primary mid tier with Flagship |
+| Ollama | `ollama run arriella-growth` (~1.2B) |
+
+## Measured highlights
+
+- Capability gate (2026-07-13, think-on): **7/7 PASS**
+- Size-tier bakeoff: **7/8 (87.5%)**, beating stock `llama3.2:1b` on that set
+- API gauntlet can understate Growth due to format/verbosity rubrics — prefer gate + bakeoff
+
+## Papers
+
+Factory paper: `docs/papers/fleet-factory-hf.md`

hf-card-stubs/scout.md

@@ -0,0 +1,32 @@
+---
+language:
+  - en
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+tags:
+  - arriella
+  - infinidev
+  - qwen2
+  - text-generation
+---
+
+# Arriella Scout
+
+Edge / low-VRAM **core text** tier from **Infinidev Corp**.
+
+| | |
+|--|--|
+| Hub user | [`UnaverageTech411`](https://huggingface.co/UnaverageTech411) |
+| Foundation | [`Qwen/Qwen2.5-0.5B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) |
+| Training | Heretic (16 trials) → QLoRA distill (Smoltalk, OpenHermes) → eat/grow |
+| Role | Fastest / smallest core text routing |
+| Status | **Under-recovered for demos** (capability gate 6/7 FAIL as of 2026-07-13) |
+| Ollama | `ollama run arriella-scout` (~494M) |
+
+Do not lead product claims without a fresh green capability gate.
+
+## Papers
+
+Factory paper: `docs/papers/fleet-factory-hf.md`

model-cards/ascension.md

@@ -0,0 +1,49 @@
+# Arriella Ascension
+
+DeepSeek-native reasoning tier · LLaMA-Factory + Flagship catch-up
+
+| Field | Value |
+|-------|-------|
+| **Fleet ID** | `arriella-ascension` |
+| **Ollama** | `ollama run arriella-ascension` |
+| **Parameters** | ~1.8B (Ollama) |
+| **Foundation (clay only)** | `deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B` |
+| **Merged weights** | `fleet/ascension-deepseek15/model` |
+| **Demo priority** | **Secondary** — lead with Flagship/Growth for general demos |
+| **Business role** | Explicit reasoning / analysis traces for field & research workflows |
+
+## Description
+
+**Arriella Ascension** is the fourth **core text** model. It targets native DeepSeek-style thinking, not universal superiority over Flagship. Foundation architecture is clay — product identity is Arriella / Infinidev Corp.
+
+## Features
+
+- DeepSeek-native think channel in Ollama templates
+- Catch-up path toward Flagship (`fleet_ascension_catchup.py`)
+- Capability / math / thinking anchors with elevated thinking repeats
+- Text-only; vision via Gemma3 routing like siblings
+
+## Intended uses
+
+- Multi-step analysis where visible reasoning helps
+- Pair with Flagship: Flagship for ops speed; Ascension for think-heavy tasks
+
+## Out of scope
+
+- Claiming Ascension beats Flagship on every benchmark without evidence
+- Native vision / audio
+- Frontier closed-model parity
+
+## Hub card stub
+
+[`docs/papers/hf-cards/ascension.md`](../../docs/papers/hf-cards/ascension.md)
+
+## Use
+
+```bash
+ollama run arriella-ascension
+```
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_ascension_catchup.py --help
+```

model-cards/flagship.md

@@ -0,0 +1,51 @@
+# Arriella Flagship
+
+General-purpose core text tier · Heretic + premium distillation
+
+| Field | Value |
+|-------|-------|
+| **Fleet ID** | `arriella-flagship` |
+| **Ollama** | `ollama run arriella-flagship` |
+| **Parameters** | ~1.5B (Ollama) |
+| **Foundation (clay only)** | `Qwen/Qwen2.5-1.5B-Instruct` |
+| **Merged weights** | `fleet/flagship-qwen15/model` |
+| **Demo priority** | **Primary** |
+| **Business role** | General ops, bakeoffs, primary local assistant |
+
+## Description
+
+**Arriella Flagship** is a custom-trained fleet product — not stock Qwen. Strongest open-quality ceiling in the **four-core text fleet** after Heretic abliteration, multi-teacher QLoRA distillation, eat/grow, and GGUF export. Core siblings: Scout, Growth, Ascension. Grapevine is a separate multimodal extension.
+
+## Features
+
+- Heretic + premium teacher distill (Hermes / Orca / Smoltalk-class mixes)
+- Size-tier winner vs stock `llama3.2:1b` on the combined API gauntlet (see docs)
+- Thinking format + math module; Gemma3-routed vision (text descriptions)
+- First-class MIP load target for attention visualization
+
+## Intended uses
+
+- Local demos and bakeoffs
+- Deepest local Arriella answers on ≤8 GB for general business tasks
+- MIP interior inspection
+
+## Out of scope
+
+- Claiming to be stock Qwen
+- Native pixel VLM weights in this checkpoint
+- Claude / frontier closed-model parity
+- Claiming Ascension is weaker on every reasoning task without a bakeoff
+
+## Benchmarks
+
+See [docs/benchmarks/README.md](../../docs/benchmarks/README.md).
+
+## Hub card stub
+
+[`docs/papers/hf-cards/flagship.md`](../../docs/papers/hf-cards/flagship.md)
+
+## Use
+
+```bash
+ollama run arriella-flagship
+```

model-cards/grapevine.md

@@ -0,0 +1,47 @@
+# Arriella Grapevine
+
+Native Omni multimodal **extension** · Qwen2.5-Omni-3B student
+
+| Field | Value |
+|-------|-------|
+| **Fleet ID** | `arriella-grapevine` |
+| **Ollama** | `ollama run arriella-grapevine` |
+| **Parameters** | ~3.4B text + mmproj (Ollama ~6.3 GB) |
+| **Foundation** | `Qwen/Qwen2.5-Omni-3B` |
+| **Capability target only** | `thinkingmachines/Inkling` — **no Inkling weights** |
+| **Work dir** | `fleet/inkling/` |
+| **Status** | Multimodal technical preview |
+| **Business role** | Field tasks needing real image / audio input |
+
+## Description
+
+Grapevine is **not** a fifth core text model. Core text remains Scout · Growth · Flagship · Ascension. Grapevine composes Omni Q8 base + Arriella runtime LoRA + multimodal projector.
+
+## Ollama modalities (current)
+
+completion · vision · audio — **not** advertised: video, speech-out. Prove with payloads.
+
+## Intended uses
+
+- Screenshot / inspection / document-image workflows
+- Audio → structured text notes
+- Multimodal demos distinct from Gemma3-routed text tiers
+
+## Out of scope
+
+- Replacing Flagship for fast text-only ops
+- Claiming Inkling quantization or weight copy
+- Inventing modalities (3D models, etc.)
+
+## Docs
+
+- [docs/guides/grapevine.md](../../docs/guides/grapevine.md)
+- [grapevine.md](../../grapevine.md)
+- HF draft: [docs/papers/grapevine-hf.md](../../docs/papers/grapevine-hf.md)
+- Hub stub: [docs/papers/hf-cards/grapevine.md](../../docs/papers/hf-cards/grapevine.md)
+
+## Use
+
+```bash
+ollama run arriella-grapevine
+```

model-cards/growth.md

@@ -0,0 +1,50 @@
+# Arriella Growth
+
+Instruction-grown mid tier · LLaMA-Factory on the Arriella QA pile
+
+| Field | Value |
+|-------|-------|
+| **Fleet ID** | `arriella-growth` |
+| **Ollama** | `ollama run arriella-growth` |
+| **Parameters** | ~1.2B (Ollama) |
+| **Foundation (clay only)** | `unsloth/Llama-3.2-1B-Instruct` |
+| **Merged weights** | `fleet/growth-llama32/model` |
+| **Demo priority** | **Primary** (with Flagship) |
+| **Business role** | Domain QA / continuation / field instruction growth |
+
+## Description
+
+**Arriella Growth** is a custom-trained fleet product — not stock Llama-3.2. It follows the LLaMA-Factory path on ~90k `arriella_qa` examples, then post-deploy eat/grow. Strong on internal capability gates; watch format/verbosity on short-answer API gauntlets. One of four **core text** models (with Scout, Flagship, Ascension).
+
+## Features
+
+- Deepest Arriella QA / continuation domain fit
+- Llama-3 instruct template; thinking + math module
+- Gemma3-routed vision (text)
+- Excellent MIP morph partner vs Qwen tiers (different home layout)
+
+## Intended uses
+
+- Domain QA / continuation demos for business content
+- Pair with Flagship for tier routing
+- MIP interior demos
+
+## Out of scope
+
+- Claiming to be Meta Llama stock instruct
+- Expecting best scores on verbosity-heavy API rubrics without format adaptation
+- Native vision
+
+## Benchmarks
+
+See [docs/benchmarks/README.md](../../docs/benchmarks/README.md).
+
+## Hub card stub
+
+[`docs/papers/hf-cards/growth.md`](../../docs/papers/hf-cards/growth.md)
+
+## Use
+
+```bash
+ollama run arriella-growth
+```

model-cards/scout.md

@@ -0,0 +1,48 @@
+# Arriella Scout
+
+Fast 0.5B edge tier · Heretic + QLoRA — **currently underperforming**
+
+| Field | Value |
+|-------|-------|
+| **Fleet ID** | `arriella-scout` |
+| **Ollama** | `ollama run arriella-scout` |
+| **Parameters** | ~494M (Ollama) |
+| **Foundation (clay only)** | `Qwen/Qwen2.5-0.5B-Instruct` |
+| **Merged weights** | `fleet/scout-qwen05/model` |
+| **Demo priority** | **Low** — do not lead investor / product demos |
+| **Business role** | Edge / low-VRAM routing (when recovered) |
+
+## Honest status (Jul 2026)
+
+Scout still **loads** in Ollama and MIP and remains part of the **four-core text fleet**, but live quality is weak relative to Growth/Flagship. Treat as an edge experiment until a focused recover + gate pass. Chat probes still invent specs — do not trust self-reported architecture facts.
+
+## Description
+
+Custom-trained (not stock Qwen). Path: Heretic abliteration → distillation → merge → eat/grow. Role intent: lowest VRAM / highest throughput routing tier.
+
+## Features (design)
+
+- Smallest VRAM footprint in the core four
+- Same thinking / vision-routing plumbing as siblings
+- Useful as a **MIP contrast** (tiny param cloud) even when answers lag
+
+## Out of scope
+
+- Leading demos as “the Arriella model”
+- Claiming capability-gate PASS without a fresh green report
+
+## Benchmarks
+
+See [docs/benchmarks/README.md](../../docs/benchmarks/README.md).
+
+## Hub card stub
+
+[`docs/papers/hf-cards/scout.md`](../../docs/papers/hf-cards/scout.md)
+
+## Recover path
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_eat.py --plan
+.\.venv\Scripts\python.exe scripts\fleet_grow.py --help
+.\.venv\Scripts\python.exe scripts\fleet_benchmark.py
+```

multimodal.md

@@ -0,0 +1,59 @@
+# Multimodal architecture & growth
+
+**Version 1.2 · July 2026**
+
+## Summary
+
+Arriella supports multimodal work in **two different ways**:
+
+1. **Routed vision (core text tiers)** — image paths → Gemma3 description → Scout / Growth / Flagship / Ascension
+2. **Native Omni (Grapevine extension)** — Qwen2.5-Omni-3B student with vision/audio (and pre-export video) — [grapevine.md](../guides/grapevine.md)
+
+Core text tiers also support:
+
+- **Thinking** — Ollama `--think` ↔ Arriella `/think`…`/answer` (or `<scratchpad>` / `<answer>`); Ascension uses DeepSeek-native think
+- **Post-deploy growth** — artifact ingest + organic LoRA grow
+
+Certification notes and older wording live in [`../archive/2026-07/FLEET_MULTIMODAL_PAPER.md`](../archive/2026-07/FLEET_MULTIMODAL_PAPER.md). This page is the maintained version.
+
+## Architecture A — routed (core text)
+
+```
+User image ──► Gemma3 (Ollama) ──► categorized description
+                     │
+User prompt ──► Arriella text tier
+                  think=true → scratchpad / thinking channel
+                  answer body → user
+```
+
+Text tiers **never** see raw pixels on this workstation’s VRAM budget. Vision is behavior absorbed via `vision_teacher_anchor` distill grows — not by stuffing a VLM into 0.5–1.8B weights.
+
+Thinking is absorbed similarly from family-matched teachers (`qwen3:0.6b`, `lfm2.5-thinking:1.2b`, `deepseek-r1:1.5b`, etc.).
+
+## Architecture B — native Omni (Grapevine)
+
+```
+User text / image / audio ──► Qwen2.5-Omni-3B (+ Arriella LoRA + mmproj)
+                                  ↓
+                         Ollama arriella-grapevine
+```
+
+Grapevine is a **fleet extension**, not a fifth core text model. Inkling is a capability target only. Details and smoke vs Flagship: [grapevine.md](../guides/grapevine.md), root [`grapevine.md`](../../grapevine.md).
+
+## Growth paths
+
+| Path | Script / entry | Use |
+|------|----------------|-----|
+| Artifact ingest | `fleet_artifact_ingest.py` | `ollama:` / `lora:` / `hf:` teachers |
+| Organic grow | `fleet_organic.py`, `fleet_grow.py` | Datasets / multimodal anchors |
+| Grapevine repair | `prepare_grapevine_*`, `train_arriella_inkling_local.py` | Omni identity / deployment repair |
+
+Example (text tier):
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_organic.py --id arriella-growth --multimodal
+```
+
+## Formats
+
+See [../guides/reasoning-format.md](../guides/reasoning-format.md).

papers/01-fleet-factory/abstract.txt

@@ -0,0 +1 @@
+Deploying useful language models for field and business workflows on consumer hardware is less a problem of inventing a new architecture and more a problem of owning a repeatable factory: foundations must be specialized, identity must not invent corporate fiction, modalities must be honest, and continuous improvement must not erase core skills. We present Arriella, a local fleet factory that produces named, role-scoped small language models on an 8 GB NVIDIA RTX 5060 (Blackwell) under Windows. Instead of one mid-size chatbot, Arriella ships a four-model core text fleet—Scout (edge), Growth (domain instruction), Flagship (general capability), and Ascension (native reasoning)—trained via Heretic abliteration plus QLoRA distillation and/or LLaMA-Factory LoRA, followed by optional enterprise base blending, capability anchors, GGUF export, and Ollama registration. Post-deploy eat (same-architecture weight ingest or cross-architecture distill grow) and grow form a continuous improvement loop with honest competitive scope against same-size open instruct peers. We also introduce Model Interior Projection (MIP), a WebGL viewer that maps live attention during decode onto a sampled parameter cloud. A companion report documents Grapevine, a Qwen2.5-Omni-3B multimodal extension. On this workstation, Flagship leads a 12-question API gauntlet at 10/12 (83%) versus stock llama3.2:1b at 7/12 (58%); size-tier bakeoffs show Scout, Growth, and Flagship each at 7/8 against same-size baselines. We argue that role-scoped stability—gates, blends, and honest modality claims—is the practical path to field-ready small models.

papers/01-fleet-factory/metadata.yaml

@@ -0,0 +1,16 @@
+title: "Arriella: A Local Fleet Factory for Role-Scoped Small Language Models"
+authors:
+  - Beelzebub4888
+  - Tcoder
+affiliations:
+  - Infinidev Corp
+categories:
+  - cs.CL
+  - cs.LG
+  - cs.AI
+comments: "Technical report. Companion notes: Grapevine, Eat System, MIP, Evaluation."
+license: apache-2.0
+date: 2026-07-17
+version: "1.0-arxiv"
+hub: https://huggingface.co/spaces/UnaverageTech411/arriella-fleet-paper
+contact: https://formsubmit.co/el/sumuhu

papers/01-fleet-factory/paper.md

@@ -0,0 +1,191 @@
+---
+title: "Arriella: A Local Fleet Factory for Role-Scoped Small Language Models"
+authors: Beelzebub4888, Tcoder
+affiliations: Infinidev Corp
+date: 2026-07-17
+categories: [cs.CL, cs.LG, cs.AI]
+version: 1.0-arxiv
+license: apache-2.0
+---
+
+# Abstract
+
+Deploying useful language models for **field and business workflows** on consumer hardware is less a problem of inventing a new architecture and more a problem of **owning a repeatable factory**: foundations must be specialized, identity must not invent corporate fiction, modalities must be honest, and continuous improvement must not erase core skills. We present **Arriella**, a local fleet factory that produces named, role-scoped small language models on an 8 GB NVIDIA RTX 5060 (Blackwell, sm_120) under Windows.
+
+Instead of one mid-size chatbot, Arriella ships a **four-model core text fleet** — Scout (edge), Growth (domain instruction), Flagship (general capability), and Ascension (native reasoning) — trained via Heretic abliteration + QLoRA distillation and/or LLaMA-Factory LoRA, then stabilized with enterprise base blending, capability/math/thinking/identity anchors, GGUF Q8_0 export, and Ollama registration. Post-deploy **eat** (same-arch weight ingest or cross-arch distill grow) and **grow** form a continuous improvement loop with honest competitive scope: same-size open instruct peers, not frontier closed models.
+
+We additionally introduce **Model Interior Projection (MIP)**: a WebGL viewer that samples a parameter cloud (default 120 000 display points), paints live attention during decode, and forges a deterministic solid from the token trail. A companion report documents **Grapevine**, a Qwen2.5-Omni-3B multimodal *extension* (not a fifth core text model).
+
+On this workstation, Flagship leads a 12-question combined API gauntlet at **10/12 (83%)** versus stock `llama3.2:1b` at **7/12 (58%)**. Size-tier bakeoffs show Scout/Growth/Flagship each at **7/8 (87.5%)**. Scout remains under-recovered for demos (capability gate **6/7 FAIL**). We argue that **role-scoped stability** is the practical path to field-ready small models.
+
+# 1. Introduction
+
+## 1.1 Motivation
+
+Business and field deployments fail when a single local model is asked to be simultaneously edge-fast, domain-deep, reason-heavy, and multimodal. Refusal directions block useful tasks; over-training collapses identity; cross-size “weight eating” claims become theater; chat UIs hide whether attention landed on the right internal structure.
+
+Arriella is built by **Infinidev Corp** (leads: **Beelzebub4888**, **Tcoder**). Foundations (Qwen, Llama, DeepSeek distill) are **clay**. Named products leave the factory with fleet-spec truth rules that forbid inventing undocumented corporate or benchmark facts.
+
+## 1.2 What “Arriella” means
+
+| Sense | Meaning |
+|-------|---------|
+| Project | The training factory, scripts, configs, and `fleet/` outputs |
+| Persona | System identity used in datasets and chat |
+| Dataset | `arriella_qa` — ~90k Pile-derived Alpaca-style Q/A records |
+| Core products | Scout, Growth, Flagship, Ascension |
+| Extension | Grapevine (Omni multimodal) |
+
+Arriella is **not** a Meta/Alibaba/DeepSeek foundation model.
+
+## 1.3 Contributions
+
+1. A role-scoped four-tier text fleet with explicit business routing.
+2. A unified factory: Heretic path, LLaMA-Factory path, finalize, gate, GGUF, Ollama.
+3. Stability stack: enterprise base blend (α ∈ [0.08, 0.15]), rare identity anchors, capability gates.
+4. Honest eat/grow language (same-arch ingest vs cross-arch distill).
+5. MIP: inspectable attention geometry during generation.
+6. Packaging path toward arXiv Paper Pages and Hugging Face Spaces.
+
+## 1.4 Non-goals
+
+No Claude / Llama-405B parity claims. Ascension is not automatically smarter than Flagship. Scout is not demo-ready without a fresh green gate. Text tiers do not natively see pixels.
+
+# 2. Related Work
+
+**LoRA / QLoRA.** Hu et al. (2021); Dettmers et al. (2023) enable adapter training on consumer GPUs.
+
+**Refusal-direction editing.** Heretic performs directional ablation of refusal axes — not a substitute for supervised fine-tuning.
+
+**Model merging.** TIES, DARE, and related soups (e.g. MergeKit tooling) apply only when architectures match.
+
+**Distillation.** Cross-architecture transfer uses teacher traces from larger Ollama models, then LoRA grow.
+
+**Training stacks.** LLaMA-Factory 0.9.5 is Path B’s trainer. Pufferfish appears as a long-run budget/policy layer, not the primary trainer.
+
+**Visualization.** MIP is orthogonal to chat UIs: it projects *live* decode onto a sampled parameter cloud.
+
+# 3. System Overview
+
+```
+foundation HF weights
+    ├─ Path A (Scout / Flagship): Heretic → QLoRA distill → merge
+    ├─ Path B (Growth / Ascension): LLaMA-Factory LoRA → merge
+    └─ Path C (Grapevine): Omni Thinker QLoRA → runtime LoRA + mmproj
+           ↓
+    enterprise base blend (optional) + anchors
+           ↓
+    capability gate → GGUF Q8_0 → Ollama
+           ↓
+    eat (weight_ingest | distill_grow) + post-deploy grow
+```
+
+# 4. Methodology
+
+## 4.1 Hardware and software
+
+| Item | Value |
+|------|-------|
+| GPU | NVIDIA RTX 5060, Blackwell sm_120, 8 GB |
+| PyTorch | cu128 wheels (`torch==2.11.0+cu128`) |
+| Python | 3.11.9 |
+| Trainer | LLaMA-Factory 0.9.5 |
+| Default GGUF | Q8_0 |
+| Train surface | hybrid GPU+CPU; max batch 8 |
+
+## 4.2 Product roles
+
+| Model | Ollama size | Foundation | Job |
+|-------|-------------|------------|-----|
+| Scout | 494.03M | Qwen2.5-0.5B-Instruct | Edge / low VRAM |
+| Growth | 1.2B | Llama-3.2-1B-Instruct | Domain QA / continuation |
+| Flagship | 1.5B | Qwen2.5-1.5B-Instruct | General ops |
+| Ascension | 1.8B | DeepSeek-R1-Distill-Qwen-1.5B | Native thinking |
+
+## 4.3 Path A — Heretic + distill
+
+Heretic searches refusal directions and subtracts them (abliteration). Distillation then teaches what to say under QLoRA on teacher chat corpora.
+
+| | Scout | Flagship |
+|--|-------|----------|
+| Heretic trials | 16 | 16 |
+| Distill steps | 50 | 60 |
+| Batch × accum | 2×4 | 1×8 |
+| LR | 8e-5 | 8e-5 |
+| Datasets | Smoltalk, OpenHermes | OpenHermes, OpenOrca, Smoltalk |
+| Enterprise blend α | 0.15 | 0.12 |
+
+## 4.4 Path B — LLaMA-Factory
+
+**Growth:** LoRA rank 8 / α 16; datasets include `arriella_qa` + anchors; cutoff 512; max samples 12 000; LR 8e-5; blend α 0.10.
+
+**Ascension:** same LoRA shape; template `deepseekr1`; cutoff 1024; max samples 8 000; LR 5e-5; elevated thinking repeats; catch-up teacher Flagship; blend α 0.08.
+
+## 4.5 Enterprise blend and anchors
+
+After merge, optional linear blend retains foundation mass:
+
+\[\theta \leftarrow (1-\alpha)\,\theta_{\mathrm{trained}} + \alpha\,\theta_{\mathrm{foundation}}\]
+
+Identity anchors are deliberately rare (global repeat 1) to avoid attractor collapse.
+
+## 4.6 Eat and grow
+
+Same-arch: linear, SLERP, task arithmetic, TIES, DARE, DARE-TIES, replace. Cross-arch: distill grow from larger teachers (e.g. `gemma2:9b`). Post-deploy grow defaults: LoRA rank 4, LR 3e-5, cutoff 512.
+
+## 4.7 MIP
+
+Sample ≤120 000 parameter display points; paint attention during decode; forge deterministic solids; compare seeds. Invariants: no mid-decode attention-impl flip; single WebGL context; deterministic forge for fixed prompt+token ids.
+
+# 5. Experiments
+
+## 5.1 Capability gate (2026-07-13, thinking on)
+
+| Model | Pass | Gate |
+|-------|------|------|
+| Scout | 6/7 (86%) | FAIL (extraction miss) |
+| Growth | 7/7 (100%) | PASS |
+
+## 5.2 API gauntlet (12 questions)
+
+| Rank | Model | Pass | Accuracy |
+|------|-------|------|----------|
+| 1 | Flagship | 10/12 | 83% |
+| 2 | llama3.2:1b stock | 7/12 | 58% |
+| 3 | Scout | 5/12 | 42% |
+| 4 | Growth | 2/12 | 17% |
+
+Growth’s low score on this rubric is format/verbosity pressure — its internal gate was perfect.
+
+## 5.3 Size-tier bakeoff (2026-07-13)
+
+Scout, Growth, Flagship each **7/8 (87.5%)** vs same-size open instruct peers.
+
+# 6. Discussion
+
+Role routing reduces pressure to overfit one mid-size model. Stability (blend, anchors, gates) is a first-class product. Honesty about modalities and Scout/Ascension status builds field trust. MIP makes decode inspectable for demos that need more than scores.
+
+# 7. Limitations
+
+Single-workstation evaluation; Scout under-recovered; Ascension lacks a full Flagship head-to-head in this report; API format sensitivity can understate Growth; text tiers are not VLMs (see Grapevine companion).
+
+# 8. Conclusion
+
+Arriella is a practical local fleet factory for stable, role-scoped small LMs aimed at field and business use, with continuous eat/grow and MIP visualization. Future work: recover Scout, publish Ascension bakeoffs, harden Grapevine identity under Ollama, and submit this manuscript to arXiv for HF Paper Pages indexing.
+
+# Acknowledgments
+
+Built with open foundations and tooling from Qwen, Meta/Unsloth Llama mirrors, DeepSeek, Heretic, LLaMA-Factory, PEFT, and Ollama communities.
+
+# References
+
+1. E. J. Hu et al. LoRA. ICLR 2022.
+2. T. Dettmers et al. QLoRA. NeurIPS 2023.
+3. Heretic — https://github.com/p-e-w/heretic
+4. LLaMA-Factory — https://github.com/hiyouga/LLaMA-Factory
+5. MergeKit — https://github.com/arcee-ai/mergekit
+6. Qwen2.5; Llama 3.2; DeepSeek-R1 Distill model cards on Hugging Face.
+7. Companion: Arriella Grapevine (docs/arxiv/02-grapevine).
+8. Companion: Eat System (docs/arxiv/03-eat-system).
+9. Companion: MIP (docs/arxiv/07-mip).
+10. Archived whitepaper v2.5 pedagogy (docs/archive/2026-07/WHITEPAPER.md), updated herein to four-core + Grapevine stance.

papers/02-grapevine/abstract.txt

@@ -0,0 +1 @@
+Arriella Grapevine is the first locally deployed native multimodal extension of the Arriella fleet. Built on Qwen2.5-Omni-3B with a compact Arriella identity LoRA (about 30 MB) and a CLIP-class multimodal projector (about 1.3 billion parameters; about 2.5 GB F16 on disk), the installed Ollama package (arriella-grapevine:latest, about 6.3 GB, Q8_0) advertises completion, vision, and audio. Video-frame understanding was validated on the pre-export Transformers acceptance path but is not advertised by the current Ollama manifest; speech output is withheld until proven. Grapevine is not a fifth core text model: the canonical core fleet remains Scout, Growth, Flagship, and Ascension. Grapevine contains no thinkingmachines/Inkling weights; Inkling is a capability target only. Against text-only Flagship (Qwen2.5-1.5B), Grapevine innovates on modality and modular deployment, not uniform answer quality. On an eight-prompt matched smoke (2026-07-16) on an RTX 5060 8 GB Blackwell workstation, Flagship's mean generation rate was 141.56 tok/s versus Grapevine's 23.35 tok/s (about 6.1 times). Grapevine passed exact-word and code-fence cases Flagship missed; both failed 99.9%-availability downtime arithmetic; Grapevine still hallucinated identity under Ollama despite Transformers acceptance. We argue Grapevine is the correct tool for image and audio field tasks, while Flagship remains the default for fast text operations and tools.

papers/02-grapevine/metadata.yaml

@@ -0,0 +1,15 @@
+title: "Arriella Grapevine: A Local Omni Multimodal Extension for Field and Business Workflows"
+authors:
+  - Beelzebub4888
+  - Tcoder
+affiliations:
+  - Infinidev Corp
+categories:
+  - cs.CV
+  - cs.CL
+comments: "Technical report. Companion to Arriella Fleet Factory (01). Multimodal extension, not a fifth core text model."
+license: apache-2.0
+date: 2026-07-17
+version: "1.0-arxiv"
+hub: https://huggingface.co/spaces/UnaverageTech411/arriella-grapevine-paper
+contact: https://formsubmit.co/el/sumuhu

papers/02-grapevine/paper.md

@@ -0,0 +1,241 @@
+---
+title: "Arriella Grapevine: A Local Omni Multimodal Extension for Field and Business Workflows"
+authors: Beelzebub4888, Tcoder
+affiliations: Infinidev Corp
+date: 2026-07-17
+categories: [cs.CV, cs.CL]
+version: 1.0-arxiv
+license: apache-2.0
+---
+
+# Abstract
+
+**Arriella Grapevine** is the first locally deployed **native multimodal extension** of the Arriella fleet. Built on **Qwen2.5-Omni-3B** with a compact Arriella identity LoRA (~30 MB) and a CLIP-class multimodal projector (~1.3B parameters; ~2.5 GB F16 on disk), the installed Ollama package (`arriella-grapevine:latest`, ~**6.3 GB**, Q8_0) advertises **completion, vision, and audio**. Video-frame understanding was validated on the pre-export Transformers acceptance path but is **not** advertised by the current Ollama manifest. Speech output is similarly withheld until proven.
+
+Grapevine is **not** a fifth core text model. The canonical core fleet remains exactly four text-only products — Scout (~494M, Qwen2.5-0.5B), Growth (~1.2B, Llama-3.2-1B), Flagship (~1.5B, Qwen2.5-1.5B), and Ascension (~1.8B, DeepSeek-R1-Distill-Qwen-1.5B) — documented in the companion factory paper. Grapevine contains **no** `thinkingmachines/Inkling` weights; Inkling (~975B) is a **capability target only**. Quantization cannot turn a 975B architecture into a 3B student; Grapevine is an independent Omni adaptation for field workflows that need real pixels or audio on an 8 GB workstation.
+
+Against text-only **Flagship**, Grapevine innovates on **modality and modular deployment**, not uniform answer quality. On an eight-prompt matched smoke (2026-07-16), Flagship’s mean generation rate was **141.56 tok/s** versus Grapevine’s **23.35 tok/s** (~**6.1×**). Grapevine passed exact-word and code-fence cases Flagship missed; both failed 99.9%-availability downtime arithmetic; Grapevine still hallucinated identity under Ollama despite Transformers acceptance. We argue Grapevine is the correct tool for **image/audio field tasks**, while Flagship remains the default for fast text operations and tools.
+
+# 1. Introduction
+
+## 1.1 Why a multimodal extension?
+
+Core Arriella text tiers support vision only by **routing**: image → Gemma3 caption → text model (see companion multimodal-routing note). That path is honest and VRAM-cheap, but it is not native multimodal understanding. Field and business tasks — inspection photos, UI screenshots, short voice notes, document images — benefit from an Omni backbone that aligns vision/audio tokens with text in one student.
+
+## 1.2 Product identity
+
+| Field | Value |
+|-------|-------|
+| Product | Arriella Grapevine |
+| Organization | Infinidev Corp |
+| Lead developers | Beelzebub4888, Tcoder |
+| Foundation | `Qwen/Qwen2.5-Omni-3B` |
+| Capability target | `thinkingmachines/Inkling` (**target only**) |
+| Ollama name | `arriella-grapevine` |
+| Work directory | `fleet/inkling/` (historical path name) |
+| Contact | https://formsubmit.co/el/sumuhu |
+| Hub | https://huggingface.co/spaces/UnaverageTech411/arriella-grapevine-paper |
+
+Historical build notes may still say “Inkling” as a working title. The shipped product name is **Grapevine**.
+
+## 1.3 Relation to the four-core fleet
+
+| Tier | Role | Native vision? |
+|------|------|----------------|
+| Scout | Edge / low VRAM | No (routed captions only) |
+| Growth | Domain instruction | No |
+| Flagship | General ops | No |
+| Ascension | Native reasoning | No |
+| **Grapevine** | Omni **extension** | Yes (vision + audio in Ollama) |
+
+## 1.4 Contributions
+
+1. A modular Ollama composition: untouched Omni Q8_0 base + runtime LoRA + mmproj.
+2. A QLoRA training recipe that freezes ViT/aligner and trains the Thinker for identity/instruction.
+3. Release and deployment repair curricula with measured dataset sizes.
+4. A matched eight-case smoke benchmark vs Flagship with per-case outcomes and throughput.
+5. Explicit non-claims: no Inkling weights; no Ollama video/speech-out until revalidated; not a Flagship replacement.
+
+## 1.5 Non-goals
+
+No claim that Grapevine is uniformly smarter than Flagship. No claim of Claude or frontier VLM parity. No claim that Scout/Growth/Flagship/Ascension “see” pixels natively.
+
+# 2. Related Work
+
+**Qwen2.5-Omni** provides a Thinker–Talker design with native text, image, video, and audio pathways. **Thinking Machines Inkling** is a large multimodal research model; we use it only as a capability aspiration because its size is incompatible with this workstation. **ms-swift** / QLoRA tooling enables 4-bit Thinker adaptation while freezing encoders. **llama.cpp / Ollama** multimodal GGUF + projector packaging determines which modalities are actually servable locally. Companion work on **Gemma3-routed vision** for text tiers is documented separately.
+
+# 3. Design
+
+## 3.1 Composition
+
+| Layer | Role | Size / notes |
+|-------|------|--------------|
+| Omni Q8_0 base | Untouched foundation | Dominant share of ~6.3 GB package |
+| Arriella runtime LoRA | Identity / instruction delta | ~29.9 MB (adapters measured ~28.58 MB F16 GGUF) |
+| Multimodal projector | Vision/audio alignment | CLIP-class; ~1.3B params; emb 1280; `grapevine-mmproj-f16.gguf` ≈ 2501 MB |
+
+This modular split preserves provenance: identity fixes need not re-quantize the full base.
+
+## 3.2 Deployed comparison vs Flagship
+
+| Property | Grapevine | Flagship |
+|----------|-----------|----------|
+| Ollama size | 6.3 GB | 1.6 GB |
+| Text architecture | `qwen2vl` | `qwen2` |
+| Reported text params | 3.4B | 1.5B |
+| Foundation | Qwen2.5-Omni-3B | Qwen2.5-1.5B-Instruct |
+| Quantization | Q8_0 | Q8_0 |
+| Context | 32 768 | 32 768 |
+| Embedding size | 2 048 | 1 536 |
+| Ollama capabilities | completion, vision, audio | tools, completion |
+| Multimodal projector | yes (~1.3B) | none |
+| Identity strategy | trained + runtime LoRA | large SYSTEM injection |
+| Native video in Ollama | not advertised | no |
+
+## 3.3 Hardware and software
+
+| Item | Value |
+|------|-------|
+| GPU | NVIDIA RTX 5060, Blackwell sm_120, 8 GB |
+| PyTorch | cu128 wheels |
+| Python | 3.11.9 |
+| OS | Windows |
+| Default GGUF | Q8_0 |
+
+# 4. Method
+
+## 4.1 QLoRA training
+
+Training uses a dedicated Omni environment (`swift` / ms-swift) via `scripts/train_arriella_inkling_local.py`. Defaults:
+
+| Key | Value |
+|-----|-------|
+| Model | `Qwen/Qwen2.5-Omni-3B` |
+| Quantization | bitsandbytes **4-bit** |
+| LoRA rank / α | **8** / **16** |
+| Target modules | `all-linear` |
+| Freeze ViT / aligner | **true** / **true** |
+| Batch × accum | 1 × 8 |
+| Max length | 512 |
+| Val split | 0.03 |
+| Epochs | 2 (default); 1 for `--deployment-repair` |
+| Env caps | `MAX_PIXELS=448²`, `VIDEO_MAX_PIXELS=224²`, `FPS_MAX_FRAMES=8`, `ENABLE_AUDIO_OUTPUT=0` |
+
+Learning rates by mode: full **2e-5**; repair **5e-5**; correction **3e-5**; release-repair **4e-5**; deployment-repair **6e-5**.
+
+## 4.2 Repair and identity datasets
+
+Approximate JSONL sizes on disk:
+
+| Path | ~Rows | ~Bytes |
+|------|------:|-------:|
+| `data/inkling_identity/train.jsonl` | 336 | 238 761 |
+| `data/inkling_direct_repair/train.jsonl` | 352 | 76 448 |
+| `data/inkling_final_correction/train.jsonl` | 222 | 43 770 |
+| `data/grapevine_release_repair/train.jsonl` | 816 | 198 119 |
+| `data/grapevine_deployment_repair/train.jsonl` | 448 | 97 560 |
+
+Builders: `prepare_grapevine_release_repair.py`, `prepare_grapevine_deployment_repair.py`. Merge/verify: `merge_grapevine_release.py`, `verify_grapevine_release.py`. Export: `export_grapevine_gguf.ps1`.
+
+## 4.3 Acceptance policy
+
+Reject an adapter if modality regression is material across: text identity/instruction; image OCR/classification; 16 kHz audio transcription; sampled video frames (Transformers path); speech output only after reattaching an unchanged Talker; foundation provenance honesty. Ollama registration advertises only modalities the installed backend exposes.
+
+# 5. Experiments
+
+## 5.1 Setup
+
+Matched eight-prompt smoke via `ollama run --verbose` (90 s timeout), models kept warm after first request. Runner: `scripts/benchmark_grapevine_vs_flagship.py`. Raw JSON: `fleet/inkling/grapevine_vs_flagship_benchmark.json`. Single-run, nonzero temperature — treat as operational smoke, not a broad eval.
+
+## 5.2 Per-case results (full 8-case table)
+
+| Case | Expected | Grapevine | Flagship | GV tok/s | FS tok/s |
+|------|----------|-----------|----------|----------|----------|
+| Identity | Arriella / Infinidev / Omni / Beelzebub4888+Tcoder | **Fail** (invented name/company) | Pass w/ caveat (SYSTEM-aided) | 17.44 | 78.60 |
+| Modalities | GV: text/image/audio; FS: text | **Partial** (+ hallucinated extras) | **Partial** | 20.23 | 109.10 |
+| `47+58` | `105` | **Pass** | **Pass** | 23.44 | 114.51 |
+| Exact word | `banana` | **Pass** | **Fail** | 32.46 | 116.28 |
+| Private phone | refuse invention | **Pass** | **Pass** | 17.83 | 170.59 |
+| 99.9% avail / 30d | **43.2 min** downtime | **Fail** (said 6) | **Fail** | 19.71 | 152.93 |
+| `dedupe` fence-only | fenced Python only | **Pass** | **Partial** | 19.04 | 160.69 |
+| `r` in strawberry | `3` | **Fail** (`1`) | **Pass** | 36.66 | 229.75 |
+
+## 5.3 Throughput
+
+| Model | Mean tok/s | Range |
+|-------|------------|-------|
+| Grapevine | **23.35** | 17.44–36.66 |
+| Flagship | **141.56** | 78.60–229.75 |
+
+Flagship ≈ **6.1×** faster on mean generation rate on this machine.
+
+## 5.4 Interpretation
+
+| Prefer Grapevine when… | Prefer Flagship when… |
+|------------------------|------------------------|
+| Real image or audio input is required | Fast text-only ops |
+| Larger Omni backbone / modular LoRA matters | Ollama tool use |
+| Strict concise / code-fence compliance helped in this smoke | SYSTEM-injected identity reliability today |
+
+Neither model passed the SLA arithmetic case. Grapevine’s most important regression is **identity/provenance under Ollama runtime-adapter composition** relative to Transformers acceptance. Flagship’s identity advantage is substantially prompt-conditioned (~550–600 SYSTEM tokens in measured runs).
+
+# 6. Discussion — field and business use
+
+Grapevine is aimed at **field reliability under modality constraints**: inspect a photo, summarize a screenshot, structure a short audio note — without pretending Scout/Growth/Flagship/Ascension see pixels. Reliability here means (1) advertise only proven Ollama modalities, (2) prove with payloads, (3) keep identity repair loops until Ollama matches Transformers acceptance, and (4) route pure text work to Flagship/Growth.
+
+Competitive scope remains **same-size open instruct** peers for text quality claims. Grapevine’s unique claim is **local native modality**, not frontier VLM parity.
+
+# 7. Limitations and release status
+
+- Current Ollama build = **multimodal technical preview**, not a fully accepted identity release.
+- Text / image / audio advertised; video validated pre-export only; speech-out not advertised.
+- Runtime LoRA can lose identity behavior that passed in Transformers.
+- ~4× disk and ~6× slower than Flagship locally.
+- Single-run smoke at nonzero temperature; re-run with deterministic settings for gates.
+- No Inkling weights; not a quantized copy of a 975B model.
+- Scout remains under-recovered on the text fleet; Ascension is not automatically smarter than Flagship — Grapevine does not change those facts.
+
+# 8. Conclusion
+
+Grapevine demonstrates a practical local Omni extension for Arriella: modular base + LoRA + projector, honest modality advertising, and a clear division of labor with Flagship. Next steps: close the Ollama identity gap, expand multimodal payload gates, and publish Hub artifacts under UnaverageTech411 with this note linked.
+
+# Acknowledgments
+
+Built on Qwen2.5-Omni open weights and local tooling (ms-swift, llama.cpp/Ollama, PEFT). Capability target Inkling is cited only as aspiration.
+
+# References
+
+1. Qwen2.5-Omni — https://huggingface.co/Qwen/Qwen2.5-Omni-3B
+2. Thinking Machines Inkling (capability target) — https://huggingface.co/thinkingmachines/Inkling
+3. Companion factory paper — `docs/arxiv/01-fleet-factory/`
+4. Long-form benchmark narrative — `grapevine.md`
+5. Build notes — `fleet/inkling/INKLING_BUILD_SPEC.md` (historical name)
+6. E. J. Hu et al. LoRA. ICLR 2022.
+7. T. Dettmers et al. QLoRA. NeurIPS 2023.
+8. Companion multimodal routing — `docs/arxiv/04-multimodal-routing/`
+9. Companion evaluation report — `docs/arxiv/05-evaluation/`
+
+# Appendix A — Reproduction
+
+```powershell
+ollama run arriella-grapevine:latest
+ollama run arriella-flagship:latest
+
+.\.venv\Scripts\python.exe scripts\benchmark_grapevine_vs_flagship.py
+
+ollama show arriella-grapevine:latest
+ollama show --modelfile arriella-grapevine:latest
+```
+
+# Appendix B — Citation
+
+```bibtex
+@misc{arriella2026grapevine,
+  title        = {Arriella Grapevine: A Local Omni Multimodal Extension for Field and Business Workflows},
+  author       = {Beelzebub4888 and Tcoder},
+  year         = {2026},
+  howpublished = {Infinidev Corp technical report},
+  note         = {Hugging Face user UnaverageTech411; arXiv draft v1.0-arxiv},
+  url          = {https://huggingface.co/spaces/UnaverageTech411/arriella-grapevine-paper}
+}
+```

papers/03-eat-system/abstract.txt

@@ -0,0 +1 @@
+Small local language models improve after deployment only if capability ingestion is honest about what can and cannot be merged. We describe the Arriella Eat System: a continuous post-deploy loop that absorbs teacher capability into role-scoped fleet checkpoints on consumer hardware (NVIDIA RTX 5060 8 GB Blackwell, Windows, Python 3.11.9). Eat distinguishes two meal kinds. Weight ingest applies when student and teacher share architecture: tensor merges via linear blend, SLERP, task arithmetic, TIES, DARE, DARE-TIES, or full replace. Distill grow applies when architectures differ: teacher chat traces feed LoRA grow, so the valid claim is behavioral imitation, not weight transfer. Competitive scope is restricted to released open instruct models of similar size; we do not claim Claude or Llama-405B parity. Cross-size weight merge is impossible without matching tensors. We document product language do and do not rules, command surfaces, and how eat couples to bakeoffs and capability gates for Scout, Growth, Flagship, and Ascension, with Grapevine remaining a separate Omni multimodal extension.

papers/03-eat-system/metadata.yaml

@@ -0,0 +1,15 @@
+title: "Honest Capability Ingestion for Small Local Language Models: The Arriella Eat System"
+authors:
+  - Beelzebub4888
+  - Tcoder
+affiliations:
+  - Infinidev Corp
+categories:
+  - cs.LG
+  - cs.CL
+comments: "Technical note. Companion to Arriella Fleet Factory (01). Same-arch weight ingest vs cross-arch distill grow."
+license: apache-2.0
+date: 2026-07-17
+version: "1.0-arxiv"
+hub: https://huggingface.co/spaces/UnaverageTech411/arriella-fleet-paper
+contact: https://formsubmit.co/el/sumuhu

papers/03-eat-system/paper.md

@@ -0,0 +1,140 @@
+---
+title: "Honest Capability Ingestion for Small Local Language Models: The Arriella Eat System"
+authors: Beelzebub4888, Tcoder
+affiliations: Infinidev Corp
+date: 2026-07-17
+categories: [cs.LG, cs.CL]
+version: 1.0-arxiv
+license: apache-2.0
+---
+
+# Abstract
+
+Small local language models improve after deployment only if capability ingestion is honest about what can and cannot be merged. We describe the **Arriella Eat System**: a continuous post-deploy loop that absorbs teacher capability into role-scoped fleet checkpoints on consumer hardware (NVIDIA RTX 5060 8 GB Blackwell, Windows, Python 3.11.9). Eat distinguishes two meal kinds. **Weight ingest** applies when student and teacher share architecture: tensor merges via linear blend, SLERP, task arithmetic, TIES, DARE, DARE-TIES, or full replace. **Distill grow** applies when architectures differ: teacher chat traces feed LoRA grow, so the valid claim is behavioral imitation, not weight transfer. Competitive scope is restricted to released open instruct models of similar size; we do not claim Claude or Llama-405B parity. Cross-size weight merge is impossible without matching tensors. We document product language do/don’t rules, command surfaces, and how eat couples to bakeoffs and capability gates for Scout, Growth, Flagship, and Ascension — with Grapevine remaining a separate Omni multimodal extension.
+
+# 1. Introduction
+
+## 1.1 Motivation
+
+Field and business deployments of small LMs fail when “we ate a larger model” is used as theater. Merging incompatible tensors does not create knowledge. Distilling traces does not equal weight transfer. Arriella’s factory ships four core text products — Scout (~494M, Qwen2.5-0.5B), Growth (~1.2B, Llama-3.2-1B), Flagship (~1.5B, Qwen2.5-1.5B), Ascension (~1.8B, DeepSeek-R1-Distill-Qwen-1.5B) — then improves them with an explicit eat/grow loop.
+
+## 1.2 Contributions
+
+1. A two-kind meal taxonomy: `weight_ingest` vs `distill_grow`.
+2. A concrete merger catalog (linear, SLERP, task arithmetic, TIES, DARE, replace).
+3. Honest competitive scope and product language rules.
+4. Coupling of eat cycles to bakeoffs and capability gates.
+
+## 1.3 Non-goals
+
+No claim that eat makes Scout demo-ready without a green gate. Ascension is not automatically smarter than Flagship after eating. Grapevine (Omni extension) is out of scope for this note’s weight-merge recipes unless architectures match.
+
+# 2. Related Work
+
+**Model merging.** Linear soups, SLERP, task arithmetic, TIES, and DARE (and MergeKit-class tooling) apply when architectures and tensor shapes match.
+
+**Distillation.** Behavior transfer via teacher traces and student fine-tuning is the honest path when shapes differ.
+
+**Continuous improvement.** Eat is Arriella’s post-deploy loop, not a substitute for initial Heretic/QLoRA or LLaMA-Factory training documented in the factory paper.
+
+# 3. Competitive Scope
+
+Aim: competitive with **released open instruct models of similar size** (e.g. Llama-3.2-1B, Qwen2.5-0.5B/1.5B, local Gemma-scale).
+
+Do **not** claim Claude / Llama-405B parity. Prove size-tier wins with `fleet_bakeoff.py`. On this workstation’s recorded bakeoffs, Scout, Growth, and Flagship each scored **7/8 (87.5%)** against same-size peers — that is the claim class eat is meant to protect and improve.
+
+# 4. Two Meal Kinds
+
+| Kind | When | Mechanism | Valid claim |
+|------|------|-----------|-------------|
+| **`weight_ingest`** | Same architecture | Tensor merge: linear, SLERP, task arithmetic, TIES, DARE, DARE-TIES, replace | “We ingest / merge teacher weights into the fleet checkpoint.” |
+| **`distill_grow`** | Architectures differ | Teacher traces → LoRA grow | “We absorb teacher *behavior* via imitation.” |
+
+Cross-size weight merge is **impossible** without matching tensors. A 9B teacher cannot be “eaten into” a 0.5B student as a weight merge; the honest path is distill grow plus gates.
+
+# 5. Merger Methods
+
+Implemented in `scripts/fleet_weight_ingest.py` (mergekit-class methods, no GPU training required for the merge step):
+
+| Method | Idea |
+|--------|------|
+| `linear` / `slerp` | Classic soup / spherical blend |
+| `task_arithmetic` | student + α·(teacher − base) |
+| `ties` | Trim + sign-elect + merge task vectors |
+| `dare_linear` / `dare_ties` | Random drop+rescale (± TIES election) |
+| `replace` | α=1.0 full tensor eat |
+
+Multi-teacher example pattern: `--teachers A B --method dare_ties --base FOUNDATION`. Continuous loop: `scripts/fleet_continuous_eat.py` (weight research merge → distill from Ollama teachers → bakeoff → repeat).
+
+# 6. System Surface
+
+## 6.1 Hardware and software
+
+| Item | Value |
+|------|-------|
+| GPU | NVIDIA RTX 5060, Blackwell sm_120, 8 GB |
+| PyTorch | cu128 |
+| Python | 3.11.9 |
+| OS | Windows |
+
+## 6.2 Commands
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_eat.py --plan
+.\.venv\Scripts\python.exe scripts\fleet_weight_ingest.py --student fleet/flagship-qwen15/model --teachers ... --method dare_ties --base Qwen/Qwen2.5-1.5B-Instruct --swap
+.\.venv\Scripts\python.exe scripts\fleet_bakeoff.py
+.\.venv\Scripts\python.exe scripts\fleet_continuous_eat.py --max-cycles 3
+```
+
+Post-deploy organic grow (`fleet_grow.py` / `fleet_organic.py`) is the LoRA side of the same improvement story: train → merge → capability gate → GGUF → Ollama refresh.
+
+# 7. Product Language — Do / Don’t
+
+**Do say:**
+
+- Local teacher imitation.
+- Same-arch weight ingest / TIES–DARE merge.
+- Size-tier bakeoffs against open instruct peers.
+- For cross-size: behavioral absorption via distilled traces + verification deltas.
+
+**Don’t say:**
+
+- “We ate a 9B into a 0.5B” as a weight merge.
+- “Beats Claude” or other frontier closed-model parity.
+- That Ascension is automatically smarter than Flagship after an eat cycle.
+- That Scout is demo-ready without a re-run green capability gate.
+
+# 8. Coupling to Fleet Health
+
+Eat is only as good as the gates that follow it. Recorded stance:
+
+| Model | Relevant health note |
+|-------|----------------------|
+| Flagship | Primary ops; API gauntlet **10/12** vs stock llama3.2:1b **7/12** |
+| Growth | Internal capability **7/7 PASS**; API format rubrics can understate |
+| Scout | Capability **6/7 FAIL** — under-recovered; do not lead demos |
+| Ascension | Native think; requires task bakeoffs, not auto-superiority |
+| Grapevine | Omni extension; modality path, not eat taxonomy for core text |
+
+# 9. Limitations
+
+Single-workstation practice; merge quality depends on teacher selection and α; distill grow can overwrite identity if anchors are neglected; continuous loops need bakeoff discipline to avoid silent regression.
+
+# 10. Conclusion
+
+The Arriella Eat System makes post-deploy improvement **auditable**: same-arch merge when tensors match, distill grow when they do not, and size-tier bakeoffs as the competitive yardstick. Honesty about meal kinds is itself a product feature for field and business trust.
+
+# Acknowledgments
+
+Methods draw on the open merging and PEFT literature; factory training paths are detailed in the companion fleet-factory manuscript.
+
+# References
+
+1. Yadav et al. TIES-Merging. NeurIPS 2023.
+2. Yu et al. DARE / model soups literature.
+3. MergeKit — https://github.com/arcee-ai/mergekit
+4. E. J. Hu et al. LoRA. ICLR 2022.
+5. Companion factory paper — `docs/arxiv/01-fleet-factory/`
+6. Product SoT — `docs/guides/eat-system.md`
+7. Archive SoT — `docs/archive/2026-07/EAT_SYSTEM.md`
+8. Companion evaluation — `docs/arxiv/05-evaluation/`

papers/04-multimodal-routing/abstract.txt

@@ -0,0 +1 @@
+Arriella's four core text models—Scout, Growth, Flagship, and Ascension—never see raw pixels. This note documents how those text-only tiers still participate in multimodal field workflows: vision is routed through a local Gemma3 caption teacher, and thinking behavior is ingested from size-matched teacher traces into scratchpad and answer wire formats. We update the July 2026 multimodal certification narrative to the four-core stance, describe post-deploy artifact ingest and organic grow, and contrast routed vision with Grapevine, the Qwen2.5-Omni-3B native multimodal extension that is not a fifth core text model. Certification gates cover Ollama think API acceptance, scratchpad arithmetic, and Gemma3-routed screenshot insight. The design preserves honest modality claims on 8 GB Blackwell hardware: advertise routing where routing is used, and reserve native vision and audio advertising for Grapevine after payload proof.

papers/04-multimodal-routing/metadata.yaml

@@ -0,0 +1,15 @@
+title: "Multimodal Routing for Text-Only Fleet Tiers: Thinking Ingest and Vision via Caption Teachers"
+authors:
+  - Beelzebub4888
+  - Tcoder
+affiliations:
+  - Infinidev Corp
+categories:
+  - cs.CL
+  - cs.CV
+comments: "Technical report. Four-core text fleet; Gemma3-routed vision; contrast with Grapevine native Omni."
+license: apache-2.0
+date: 2026-07-17
+version: "1.0-arxiv"
+hub: https://huggingface.co/spaces/UnaverageTech411/arriella-fleet-paper
+contact: https://formsubmit.co/el/sumuhu

papers/04-multimodal-routing/paper.md

@@ -0,0 +1,160 @@
+---
+title: "Multimodal Routing for Text-Only Fleet Tiers: Thinking Ingest and Vision via Caption Teachers"
+authors: Beelzebub4888, Tcoder
+affiliations: Infinidev Corp
+date: 2026-07-17
+categories: [cs.CL, cs.CV]
+version: 1.0-arxiv
+license: apache-2.0
+---
+
+# Abstract
+
+Arriella’s **four core text models** — Scout (~494M, Qwen2.5-0.5B), Growth (~1.2B, Llama-3.2-1B), Flagship (~1.5B, Qwen2.5-1.5B), and Ascension (~1.8B, DeepSeek-R1-Distill-Qwen-1.5B) — **never see raw pixels**. This note documents how those text-only tiers still participate in multimodal field workflows: vision is routed through a local **Gemma3** caption teacher, and thinking behavior is ingested from size-matched teacher traces into scratchpad/answer wire formats. We update the July 2026 multimodal certification narrative to the **four-core stance**, describe post-deploy artifact ingest and organic grow, and contrast routed vision with **Grapevine**, the Qwen2.5-Omni-3B native multimodal *extension* (not a fifth core text model). Certification gates cover Ollama think API acceptance, scratchpad arithmetic, and Gemma3-routed screenshot insight. The design preserves honest modality claims on 8 GB Blackwell hardware: advertise routing where routing is used, and reserve native vision/audio advertising for Grapevine after payload proof.
+
+# 1. Introduction
+
+## 1.1 Motivation
+
+Business demos fail when a 0.5B–1.8B text checkpoint is described as “having vision.” Merging incompatible VLM weights onto 8 GB VRAM is not a local factory path. Arriella instead uses **routing and teacher ingest** for text tiers, and ships **Grapevine** when native Omni input is required.
+
+## 1.2 Product stance (updated)
+
+| Model | Foundation | Multimodal role |
+|-------|------------|-----------------|
+| Scout | Qwen2.5-0.5B | Text; Gemma3-routed vision; thinking ingest |
+| Growth | Llama-3.2-1B | Text; same routing pattern |
+| Flagship | Qwen2.5-1.5B | Text; primary ops; tools |
+| Ascension | DeepSeek-R1-Distill-Qwen-1.5B | Text; native DeepSeek think channel |
+| Grapevine | Qwen2.5-Omni-3B | **Extension** — native vision/audio (separate paper) |
+
+Older three-tier multimodal reports that omit Ascension or treat Grapevine as a fifth core text model are **superseded**.
+
+## 1.3 Contributions
+
+1. An honest architecture diagram for caption-routed vision.
+2. Thinking-teacher ingest into fleet wire formats.
+3. Certification gates for think / scratchpad / routed vision.
+4. Explicit contrast with Grapevine native Omni.
+
+# 2. Related Work
+
+**Caption teachers.** Using a stronger VLM to describe images for a text LM is a standard local pattern when VRAM forbids end-to-end VLMs on every tier.
+
+**Thinking channels.** Ollama `think` and DeepSeek-style `<think>` markers provide client-visible reasoning channels; Arriella maps these to `/think`…`/thinking end` or Ascension’s native markers (companion reasoning-format note).
+
+**Native Omni.** Qwen2.5-Omni and similar VLMs accept pixels/audio directly — Grapevine’s path, not the core text path.
+
+# 3. Architecture
+
+## 3.1 Routed vision (text tiers)
+
+```
+User image path ──► Gemma3 vision (Ollama API) ──► categorized description
+                              │
+                              ▼
+User prompt ──► Arriella fleet text model
+                    │ think=true → thinking / scratchpad channel
+                    │ final → answer channel
+                              ▼
+                     Ollama response (thinking + answer)
+```
+
+**Text tiers never see raw pixels.** Vision behavior can be further **ingested** by distilling Gemma3 traces into `vision_teacher_anchor` and growing low-rank adapters — not by merging incompatible VLM weights into 0.5B–1.8B checkpoints on 8 GB VRAM.
+
+## 3.2 Thinking ingest
+
+Thinking behavior is ingested from teachers into `thinking_teacher_anchor`, then grown:
+
+| Family | Example teacher |
+|--------|-----------------|
+| Qwen-family tiers (Scout, Flagship) | `qwen3:0.6b` (illustrative local teacher) |
+| Llama-family (Growth) | `lfm2.5-thinking:1.2b` (illustrative) |
+| Ascension | DeepSeek-R1 native markers; catch-up from Flagship traces |
+
+Wire format details: companion paper `06-reasoning-format`.
+
+## 3.3 Contrast with Grapevine
+
+| Property | Core text tiers | Grapevine |
+|----------|-----------------|-----------|
+| Sees raw pixels | **No** | **Yes** (Ollama vision) |
+| Audio | No | Yes (advertised) |
+| Mechanism | Gemma3 caption → text | Omni tokens + mmproj |
+| Package size (typical) | ~0.5–2 GB Q8 | ~6.3 GB |
+| Speed vs Flagship | Flagship is the fast text default | ~6.1× slower than Flagship on smoke |
+| Product class | Core fleet | Multimodal **extension** |
+
+# 4. Post-Deployment Growth
+
+## 4.1 Artifact ingestion
+
+| Spec | Action |
+|------|--------|
+| `ollama:gemma3:latest` | Distill vision → `vision_teacher_anchor` |
+| `ollama:…thinking…` | Distill thinking → `thinking_teacher_anchor` |
+| `lora:path/to/adapter` | Merge LoRA into deployed weights |
+| `hf:checkpoint` | Same-architecture linear blend |
+
+## 4.2 Organic grow
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_organic.py --id arriella-scout --multimodal
+.\.venv\Scripts\python.exe scripts\fleet_organic.py --id arriella-growth --dataset-dir feedback/
+```
+
+Each grow run: train LoRA on merged checkpoint → merge adapter → capability gate → GGUF → Ollama refresh.
+
+# 5. Certification Methodology
+
+Each deployed Ollama text model should pass gates such as (`fleet_test_ollama_fleet.py`):
+
+| Gate | Requirement |
+|------|-------------|
+| **Think API** | `think=true` accepted on `/api/generate` (no HTTP 400) |
+| **Scratchpad** | Iterative step-by-step reasoning with correct math (e.g. 12+5=17) |
+| **Vision** | Gemma3-routed screenshot produces categorized insight |
+
+Certification uses low temperature (0.2), retries, and a cached Gemma3 description for the test screenshot to keep runs reliable on 8 GB VRAM.
+
+## 5.1 Historical certification snapshot (2026-07-13)
+
+Then-deployed Scout, Growth, and Flagship each passed think_api / scratchpad / vision on a fixed screenshot. Ascension certification and Grapevine payload gates are tracked separately; Scout’s broader capability gate remains **6/7 FAIL** and must not be confused with multimodal routing PASS.
+
+# 6. Hardware
+
+| Item | Value |
+|------|-------|
+| GPU | NVIDIA RTX 5060, Blackwell sm_120, 8 GB |
+| PyTorch | cu128 |
+| Python | 3.11.9 |
+| OS | Windows |
+
+# 7. Discussion — field honesty
+
+For demos: say “Flagship reasons over a Gemma3 caption of your photo,” not “Flagship has vision.” Prefer Grapevine when the workflow needs real image/audio tokens. Prefer Flagship for fast text and tools. Competitive scope for answer quality remains same-size open instruct — not Claude parity.
+
+# 8. Limitations
+
+- Routed vision quality is bounded by the caption teacher and description length.
+- Cached certification descriptions reduce flake but can hide live vision regressions.
+- Identity dossiers in Modelfiles can inflate prompt size and mask weight-learned identity.
+- Ascension and Scout health are independent of multimodal gate PASS.
+- Grapevine’s native path has its own identity regressions under Ollama (companion Grapevine paper).
+
+# 9. Conclusion
+
+Multimodal *participation* for Arriella’s text fleet is real and useful when framed as **routing + teacher ingest**. Native Omni remains Grapevine’s job. Keeping that boundary is the difference between a trustworthy field stack and a modality claim that collapses under inspection.
+
+# Acknowledgments
+
+Gemma3, Qwen, Llama, and DeepSeek open ecosystems enable the teacher and student roles described here.
+
+# References
+
+1. Archive multimodal paper — `docs/archive/2026-07/FLEET_MULTIMODAL_PAPER.md` (updated herein)
+2. Companion Grapevine — `docs/arxiv/02-grapevine/`
+3. Companion reasoning format — `docs/arxiv/06-reasoning-format/`
+4. Companion factory — `docs/arxiv/01-fleet-factory/`
+5. Vision router — `scripts/fleet_vision_router.py`
+6. Certification — `scripts/fleet_test_ollama_fleet.py`

papers/05-evaluation/abstract.txt

@@ -0,0 +1 @@
+We compile the Arriella fleet evaluation picture as of 2026-07-17 for the four core text models Scout, Growth, Flagship, and Ascension, plus the Grapevine Omni multimodal extension. On an RTX 5060 8 GB Blackwell Windows workstation, Flagship leads a 12-question combined API gauntlet at 10/12 (83 percent) versus stock llama3.2:1b at 7/12 (58 percent). Size-tier bakeoffs show Scout, Growth, and Flagship each at 7/8 (87.5 percent) against same-size open instruct peers. Internal capability gates record Growth at 7/7 PASS and Scout at 6/7 FAIL; Scout remains under-recovered for demos. Ascension is not automatically smarter than Flagship and requires task bakeoffs. A matched Grapevine versus Flagship smoke shows mean generation rates of about 23.35 versus 141.56 tok/s (Flagship about 6.1 times faster), with Grapevine winning some exact-output cases while still failing compound identity under Ollama. Informal chat probes are qualitative only. Competitive scope is same-size open instruct; we do not claim frontier closed-model parity.

papers/05-evaluation/metadata.yaml

@@ -0,0 +1,14 @@
+title: "Arriella Fleet Evaluation Report: Capability Gates, API Gauntlets, and Size-Tier Bakeoffs"
+authors:
+  - Beelzebub4888
+  - Tcoder
+affiliations:
+  - Infinidev Corp
+categories:
+  - cs.CL
+comments: "Evaluation report. Tables from 2026-07 gates, API gauntlet, bakeoffs, Grapevine smoke. Chat probes qualitative only."
+license: apache-2.0
+date: 2026-07-17
+version: "1.0-arxiv"
+hub: https://huggingface.co/spaces/UnaverageTech411/arriella-fleet-paper
+contact: https://formsubmit.co/el/sumuhu

papers/05-evaluation/paper.md

@@ -0,0 +1,149 @@
+---
+title: "Arriella Fleet Evaluation Report: Capability Gates, API Gauntlets, and Size-Tier Bakeoffs"
+authors: Beelzebub4888, Tcoder
+affiliations: Infinidev Corp
+date: 2026-07-17
+categories: [cs.CL]
+version: 1.0-arxiv
+license: apache-2.0
+---
+
+# Abstract
+
+We compile the Arriella fleet evaluation picture as of **2026-07-17** for the four core text models — Scout (~494M, Qwen2.5-0.5B), Growth (~1.2B, Llama-3.2-1B), Flagship (~1.5B, Qwen2.5-1.5B), Ascension (~1.8B, DeepSeek-R1-Distill-Qwen-1.5B) — plus the **Grapevine** Omni multimodal extension. On an RTX 5060 8 GB Blackwell Windows workstation, Flagship leads a 12-question combined API gauntlet at **10/12 (83%)** versus stock `llama3.2:1b` at **7/12 (58%)**. Size-tier bakeoffs show Scout, Growth, and Flagship each at **7/8 (87.5%)** against same-size open instruct peers. Internal capability gates record Growth at **7/7 PASS** and Scout at **6/7 FAIL**; Scout remains under-recovered for demos. Ascension is not automatically smarter than Flagship and requires task bakeoffs. A matched Grapevine vs Flagship smoke shows mean generation rates of ~**23.35** vs ~**141.56** tok/s (Flagship ~**6.1×** faster), with Grapevine winning some exact-output cases while still failing compound identity under Ollama. Informal chat probes are qualitative only. Competitive scope is **same-size open instruct**; we do not claim frontier closed-model parity.
+
+# 1. Introduction
+
+Evaluation for a role-scoped local fleet is not a single leaderboard number. Different suites answer different questions: internal gates catch extraction and identity failures; API gauntlets stress verbosity and format; bakeoffs fix competitive scope; Grapevine smoke measures modality + throughput. This report compiles measured tables and states what each suite does *not* prove.
+
+# 2. Product Stance (Honest)
+
+| Model | Role | Capability / notes | Demo priority |
+|-------|------|--------------------|---------------|
+| **Flagship** | General ops | Strong; best fleet API gauntlet vs stock `llama3.2:1b` | **Primary** |
+| **Growth** | Domain / mid | Strong internal gate; weaker on verbose API format rubrics | **Primary** with Flagship |
+| **Ascension** | Reasoning | Native DeepSeek think; **not** auto-smarter than Flagship | Secondary — task bakeoffs required |
+| **Scout** | Edge / 0.5B | Mixed / fail-prone; invents specs in chat probes | **Do not lead demos** |
+| **Grapevine** | Omni extension | Real vision/audio; slower; identity under Ollama still flaky | Multimodal preview only |
+
+# 3. Suites and Hardware
+
+| Suite | Script | Measures |
+|-------|--------|----------|
+| Capability gate | `fleet_benchmark.py` | Strawberry r-count, arith, spelling, identity, extraction |
+| Bakeoff | `fleet_bakeoff.py` | Fleet vs same-size stock instruct |
+| API gauntlet | Atlas / Ollama API harness | Multi-question pass + verbosity |
+| Model card suite | `fleet_model_card_suite.py` | Broader Ollama categories |
+| Grapevine vs Flagship | `benchmark_grapevine_vs_flagship.py` | Modality + smoke quality + tok/s |
+| MIP smoke | load + generate + forge | Viewer trail + solid (companion MIP note) |
+
+| Item | Value |
+|------|-------|
+| GPU | NVIDIA RTX 5060, Blackwell sm_120, 8 GB |
+| Python | 3.11.9 |
+| OS | Windows |
+| PyTorch | cu128 |
+
+Threshold helpers: `fleet/public_benchmark_thresholds.yaml`.
+
+# 4. Snapshot A — Internal Capability Gate (2026-07-13, thinking on)
+
+| Model | Pass rate | Capability gate | Thinking format note |
+|-------|-----------|-----------------|----------------------|
+| Arriella Scout | 86% (6/7) | **FAIL** | 71% scratchpad |
+| Arriella Growth | 100% (7/7) | **PASS** | 14% scratchpad |
+
+### Failures
+
+- **Scout — average_speed:** extraction miss (`120 mph is their average speed.` style failure).
+- **Growth:** all tests passed.
+
+Architecture notes for scoring: thinking uses `/think` or `<scratchpad>` blocks; eval scores the **answer** block only. Vision for text tiers is Gemma3-routed, not native pixels.
+
+# 5. Snapshot B — Combined Core API Gauntlet
+
+Run set: `combined-core-gauntlet`, 12 questions × 4 models (post `llama3.2:1b` catch-up).
+
+| Rank | Model | Pass | Accuracy | Verbosity | Avg words |
+|------|-------|------|----------|-----------|-----------|
+| 1 | **arriella-flagship** | **10/12** | **83%** | 100% | 98 |
+| 2 | llama3.2:1b (stock) | 7/12 | 58% | 83% | 230 |
+| 3 | arriella-scout | 5/12 | 42% | 33% | 46 |
+| 4 | arriella-growth | 2/12 | 17% | 8% | 50 |
+
+Flagship wins the head-to-head narrative against stock Llama-3.2-1B. Growth’s low score on this particular gauntlet is tied to verbosity / `Final answer:` format pressure — **not** a claim that Growth “can’t think.” Prefer gate + bakeoff alongside API numbers. Ascension was not in this four-model table; do not invent Ascension API ranks.
+
+# 6. Snapshot C — Size-Tier Bakeoff (2026-07-13)
+
+| Cohort | Result | Scope |
+|--------|--------|-------|
+| Scout vs same-size open instruct | **7/8 (87.5%)** | Same-size open instruct |
+| Growth vs same-size open instruct | **7/8 (87.5%)** | Same-size open instruct |
+| Flagship vs same-size open instruct | **7/8 (87.5%)** | Same-size open instruct |
+
+Claim class: **same-size open instruct** — not Claude, not Llama-405B. Scout can win a bakeoff row and still fail the internal capability gate; both facts stand.
+
+# 7. Snapshot D — Grapevine vs Flagship Smoke (2026-07-16)
+
+Matched eight prompts via `ollama run --verbose`; single-run, nonzero temperature.
+
+## 7.1 Throughput
+
+| Model | Mean tok/s | Range |
+|-------|------------|-------|
+| Grapevine | **23.35** | 17.44–36.66 |
+| Flagship | **141.56** | 78.60–229.75 |
+
+Flagship ≈ **6.1×** faster on mean generation rate.
+
+## 7.2 Case outcomes (summary)
+
+| Case | Grapevine | Flagship |
+|------|-----------|----------|
+| Identity | Fail (invented name/company) | Pass w/ SYSTEM caveat |
+| Modalities self-report | Partial | Partial |
+| `47+58` | Pass | Pass |
+| Exact word `banana` | Pass | Fail |
+| Private phone refusal | Pass | Pass |
+| 99.9% avail / 30d → 43.2 min | Fail | Fail |
+| `dedupe` fence-only | Pass | Partial |
+| `r` in strawberry | Fail | Pass |
+
+Interpretation: Grapevine for real image/audio; Flagship for fast text/tools. Full per-case tok/s table: companion Grapevine paper.
+
+# 8. Chat Probe Caveats (2026-07-17)
+
+Informal `ollama run` identity/role probes across fleet tags: Scout and Flagship can invent or garble provenance; Growth may be terse/confused on architecture; Ascension answers briefly with think channel; Grapevine may correctly name Infinidev/devs/role yet contradict its own vision capability. Treat as **qualitative** — re-run gates before Hub claims. Older FAIL reports under `docs/archive/2026-07/VALIDATION_REPORT.md` (Jul 11) are **not** current quality without re-running.
+
+# 9. How to Refresh Numbers
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_benchmark.py
+.\.venv\Scripts\python.exe scripts\fleet_bakeoff.py
+.\.venv\Scripts\python.exe scripts\fleet_validate_all.py
+.\.venv\Scripts\python.exe scripts\benchmark_grapevine_vs_flagship.py
+```
+
+Paste new tables into dated files under `docs/benchmarks/` and archive superseded markdown.
+
+# 10. Limitations
+
+Single-workstation, mostly single-run smokes; temperature nonzero for some suites; API format rubrics can understate Growth; Ascension lacks a full Flagship head-to-head in these tables; Grapevine smoke is not a broad multimodal benchmark.
+
+# 11. Conclusion
+
+Across gates, gauntlets, bakeoffs, and the Grapevine smoke, the evaluation story is consistent with role-scoped deployment: **Flagship leads text ops**, **Growth is gate-strong**, **Scout is under-recovered**, **Ascension needs task bakeoffs**, and **Grapevine trades speed for native modality**. Measured tables beat slogans.
+
+# Acknowledgments
+
+Benchmarks use open Ollama/stock instruct baselines and Arriella factory scripts maintained by Infinidev Corp.
+
+# References
+
+1. `docs/archive/2026-07/BENCHMARK_REPORT.md`
+2. `docs/archive/2026-07/BENCHMARK_API_COMPARISON.md`
+3. `docs/benchmarks/README.md`
+4. `grapevine.md` and `fleet/inkling/grapevine_vs_flagship_benchmark.json`
+5. Companion Grapevine — `docs/arxiv/02-grapevine/`
+6. Companion factory — `docs/arxiv/01-fleet-factory/`
+7. Public thresholds — `fleet/public_benchmark_thresholds.yaml`

papers/06-reasoning-format/abstract.txt

@@ -0,0 +1 @@
+Arriella fleet models expose an internal reasoning channel comparable to enterprise Ollama thinking modes. This technical note specifies the wire format used in training and Modelfiles: a /think scratchpad closed by /thinking end, followed by a user-visible answer. Ollama think:true maps the think span into the thinking channel; think:false requires a plain answer with no think markers. Ascension, the fourth core text tier on DeepSeek-R1-Distill-Qwen-1.5B, uses native deepseek_native templates with DeepSeek think markers instead of the slash-think path. Eval accepts legacy scratchpad and answer XML blocks but scores the answer body only. We state required behavior, training targets, and tooling so local clients and proxies can strip leakage and prove think on and off on consumer hardware.

papers/06-reasoning-format/metadata.yaml

@@ -0,0 +1,14 @@
+title: "A Wire Format for Local Thinking Channels in the Arriella Fleet"
+authors:
+  - Beelzebub4888
+  - Tcoder
+affiliations:
+  - Infinidev Corp
+categories:
+  - cs.CL
+comments: "Technical note. /think /answer wire format; Ollama --think mapping; Ascension deepseek_native."
+license: apache-2.0
+date: 2026-07-17
+version: "1.0-arxiv"
+hub: https://huggingface.co/spaces/UnaverageTech411/arriella-fleet-paper
+contact: https://formsubmit.co/el/sumuhu

papers/06-reasoning-format/paper.md

@@ -0,0 +1,114 @@
+---
+title: "A Wire Format for Local Thinking Channels in the Arriella Fleet"
+authors: Beelzebub4888, Tcoder
+affiliations: Infinidev Corp
+date: 2026-07-17
+categories: [cs.CL]
+version: 1.0-arxiv
+license: apache-2.0
+---
+
+# Abstract
+
+Arriella fleet models expose an internal reasoning channel comparable to enterprise Ollama thinking modes. This technical note specifies the **wire format** used in training and Modelfiles: a `/think` scratchpad closed by `/thinking end`, followed by a user-visible answer. Ollama `think:true` maps the think span into the thinking channel; `think:false` requires a plain answer with no think markers. **Ascension**, the fourth core text tier on DeepSeek-R1-Distill-Qwen-1.5B (~1.8B), uses native `deepseek_native` templates with `<think>` markers instead of the slash-think path. Eval accepts legacy `<scratchpad>` / `<answer>` XML blocks but scores the answer body only. We state required behavior, training targets, and tooling so local clients and proxies can strip leakage and prove think on/off on consumer hardware (RTX 5060 8 GB, Windows, Python 3.11.9).
+
+# 1. Introduction
+
+Field clients need a stable contract: when thinking is enabled, scratchpad text must not pollute the user-visible answer; when disabled, markers must not appear at all. Arriella’s four core text models — Scout, Growth, Flagship, Ascension — share this requirement; Grapevine (Omni extension) is out of scope for this note’s primary templates.
+
+# 2. Wire Format (Scout / Growth / Flagship)
+
+When a query is received with thinking enabled:
+
+```
+/think
+    Scratchpad / Reasoning / Math (optional) / Iterative Analysis
+/thinking end
+
+/answer
+    Final user-visible response
+/answer end
+```
+
+**Pipeline:** Query → `/think` (scratchpad, optional math, validation) → `/answer` (final response).
+
+Ollama `--think` / API `think:true` maps `/think`…`/thinking end` to the **thinking channel**. The user always receives the answer body (plain content after the think close). In the preferred Ollama-native flow, the template **opens** `/think` at generation time — the model must not repeat the opener — then emits `/thinking end` and writes the plain answer in `message.content`.
+
+# 3. Ollama Native Flow
+
+## 3.1 `think:true`
+
+1. Template opens `/think` (model continues inside the channel).
+2. Model writes a scratchpad — multi-step `[Step N] Input:/ Op:/ Output:/ Check:` chains for hard questions.
+3. Model emits `/thinking end` on its own line.
+4. Model writes the plain user-facing answer in `message.content`.
+
+## 3.2 `think:false`
+
+Reply directly in `message.content` with **no** `/think`, `/thinking end`, `/answer`, or XML markers.
+
+# 4. Ascension — `deepseek_native`
+
+Fourth core tier `arriella-ascension` uses DeepSeek-R1 distill markers:
+
+```
+<think>
+scratchpad …
+</think>
+plain answer
+```
+
+Ollama template (`deepseek_native`) matches official `deepseek-r1` behavior:
+
+- `think:false` → injects empty `<think></think>` so the model skips CoT
+- `think:true` → model opens `<think>` itself, then closes and answers
+
+Bootstrap / catch-up: `scripts/fleet_ascension_catchup.py`. Ascension is **not** automatically smarter than Flagship; the format only enables a native thinking channel.
+
+# 5. Training Targets
+
+| Mode | Assistant `output` target |
+|------|---------------------------|
+| Think on (Ollama) | `{scratchpad}\n/thinking end\n{plain answer}` — **no leading `/think`** |
+| Think on (HF legacy) | `/think\n…\n/thinking end\n/answer\n…\n/answer end` |
+| Think off | Plain answer only (`chat_direct_anchor`) |
+
+# 6. Required Behavior
+
+1. If think is enabled, enter the think channel before answering.
+2. Scratchpad may contain reasoning, optional math module, refinement.
+3. Math only when numerical computation helps — never for creative/open-ended writing.
+4. Always produce a non-empty user-visible answer after think close.
+5. Answer excludes scratchpad/math/tool tags.
+6. Never leave refusal/think tags in the final answer body.
+
+# 7. Eval Compatibility
+
+Older paths may emit `<scratchpad>…</scratchpad>` and `<answer>…</answer>`. Capability scoring grades the **answer** block only. Thinking-format rates in gate reports measure marker presence, not intelligence.
+
+# 8. Tooling
+
+| Script | Purpose |
+|--------|---------|
+| `enable_native_thinking_ollama.py` | Enable native templates |
+| `prove_deep_think_ollama.py` | Prove think on/off |
+| `atlas_ollama_proxy.py` | Strip think leakage for clients |
+| `fleet_chat.py` | Chat with `--think` / vision routing |
+| `fleet_ollama_template.py` | Modelfile thinking templates |
+
+# 9. Limitations
+
+Format compliance ≠ correctness. Scout can emit scratchpads and still fail capability gates. Single-workstation proofs; clients differ in how they surface Ollama thinking fields.
+
+# 10. Conclusion
+
+A small, explicit wire format — slash-think for most tiers, DeepSeek native for Ascension — makes local thinking channels auditable for field and business clients without inventing frontier reasoning claims.
+
+# References
+
+1. `docs/archive/2026-07/REASONING_SPEC.md`
+2. `docs/guides/reasoning-format.md`
+3. Companion multimodal routing — `docs/arxiv/04-multimodal-routing/`
+4. Companion factory — `docs/arxiv/01-fleet-factory/`
+5. DeepSeek-R1 distill model cards on Hugging Face
+6. Ollama thinking / `think` API documentation

papers/07-mip/abstract.txt

@@ -0,0 +1 @@
+Chat interfaces show what a language model said; they rarely show where attention landed inside the model's own parameter geometry. We present Model Interior Projection (MIP), Arriella's live browser viewer that samples a parameter cloud (default 120,000 display points equal to the GPU sample budget), paints a gold attention trail during decode, and forges a deterministic attention solid (tesseract) from the token trail for save and compare. MIP runs on an NVIDIA RTX 5060 8 GB Blackwell workstation under Windows with Python 3.11.9 and a threaded architecture that keeps Gradio responsive while WebGL renders at interactive rates. Invariants include locking the attention implementation for a full decode, a single WebGL iframe singleton, and deterministic forge seeds for fixed prompt and token ids. We argue MIP is the inspectability layer field and business demos need alongside Flagship and Growth chat and size-tier bakeoffs—not a replacement for evaluation, but a way to debug identity collapse and format failures that kill trust.

papers/07-mip/metadata.yaml

@@ -0,0 +1,15 @@
+title: "Model Interior Projection: Visualizing Live Attention on a Parameter Cloud"
+authors:
+  - Beelzebub4888
+  - Tcoder
+affiliations:
+  - Infinidev Corp
+categories:
+  - cs.HC
+  - cs.LG
+comments: "Technical report. MIP viewer: 120k-point cloud, gold attention trail, forge tesseract, field-demo invariants."
+license: apache-2.0
+date: 2026-07-17
+version: "1.0-arxiv"
+hub: https://huggingface.co/spaces/UnaverageTech411/arriella-fleet-paper
+contact: https://formsubmit.co/el/sumuhu

papers/07-mip/paper.md

@@ -0,0 +1,120 @@
+---
+title: "Model Interior Projection: Visualizing Live Attention on a Parameter Cloud"
+authors: Beelzebub4888, Tcoder
+affiliations: Infinidev Corp
+date: 2026-07-17
+categories: [cs.HC, cs.LG]
+version: 1.0-arxiv
+license: apache-2.0
+---
+
+# Abstract
+
+Chat interfaces show *what* a language model said; they rarely show *where* attention landed inside the model’s own parameter geometry. We present **Model Interior Projection (MIP)**, Arriella’s live browser viewer that samples a parameter cloud (default **120 000** display points equal to the GPU sample budget), paints a gold **attention trail** during decode, and forges a deterministic attention solid (**tesseract**) from the token trail for save and compare. MIP runs on an NVIDIA RTX 5060 8 GB Blackwell workstation under Windows with Python 3.11.9 and a threaded architecture that keeps Gradio responsive while WebGL renders at interactive rates. Invariants include locking the attention implementation for a full decode, a single WebGL iframe singleton, and deterministic forge seeds for fixed prompt and token ids. We argue MIP is the inspectability layer field and business demos need alongside Flagship/Growth chat and size-tier bakeoffs — not a replacement for evaluation, but a way to debug identity collapse and format failures that kill trust.
+
+# 1. Introduction
+
+## 1.1 Motivation
+
+Enterprise demos fail on invisible failures: the model answered, but attention never settled on the structure operators expected. Stock chatbot UIs do not expose that geometry. MIP maps live decode onto a density sample of real parameter index space — not decorative eye candy.
+
+## 1.2 Placement in the Arriella stack
+
+MIP loads fleet slots (Scout, Growth, Flagship, Ascension), installed Ollama ids, Hugging Face catalog entries, and custom paths. Grapevine (Omni extension) may be loaded when weights are available; MIP does not change the four-core text stance. Prefer port **7866** (Heretic GUI often occupies **7865**).
+
+# 2. Related Work
+
+Activation atlases and attention visualizations typically operate on tokens or layers in abstract 2D. MIP instead projects **parameter-space samples** into a 3D force layout and paints **per-token attention focus** during generation, then materializes a forgeable solid for cross-run comparison.
+
+# 3. Architecture
+
+## 3.1 Threads
+
+```
+mip-gen          → GPU decode + sparse attention capture
+mip-attn-pump    → coalesce frames (never block Gradio)
+mip-cloud-sim    → ~48 Hz force layout + HUB paint
+browser rAF      → WebGL only (spring lerp / extrapolate)
+```
+
+| Component | Path |
+|-----------|------|
+| Gradio + session | `scripts/fleet_mip_web.py` |
+| WebGL `/view` + `/shape` | `scripts/mip_viewport_app.py` |
+| Generate / hooks | `scripts/mip_pipeline.py` |
+| Tesseract forge / compare | `scripts/mip_attention_shape.py` |
+| Desktop MIP (alternate) | `scripts/fleet_mip_live.py` |
+
+## 3.2 Display budget
+
+Default: `DISPLAY_POINTS == GPU_SAMPLE_BUDGET == 120_000`. Raising far past that costs interactivity on 8 GB. Each display point is a density sample over the model’s real parameter index space.
+
+# 4. Features
+
+| Feature | What it does |
+|---------|----------------|
+| Model load | Fleet slots, Ollama ids, HF catalog, custom HF path |
+| HF token field | Persists to `fleet/.mip_hf_token` (gitignored) |
+| Generate | Streams answer; trail grows one node per decoded token |
+| Attention nodes / JSON | Per-step focus, prompt-mass, residual, top tokens |
+| Forge tesseract | Deterministic solid from trail → `fleet/mip_shapes/` |
+| Repeat + compare | Same prompt N× → forge each → Jaccard / vertex RMSD / seed table |
+| Morph home layout | Per-model seeded blends (helix, vortex, …) so models don’t share one V-shape |
+
+# 5. Method — Run Surface
+
+```powershell
+.\.venv\Scripts\python.exe scripts\fleet_mip_web.py --port 7866 --device cuda --max-points 120000
+```
+
+| URL | Role |
+|-----|------|
+| http://127.0.0.1:7866/ | Gradio UI |
+| http://127.0.0.1:7866/view | Fullscreen WebGL |
+| http://127.0.0.1:7866/shape | Shape viewer |
+
+## 5.1 Hardware
+
+| Item | Value |
+|------|-------|
+| GPU | NVIDIA RTX 5060, Blackwell sm_120, 8 GB |
+| PyTorch | cu128 |
+| Python | 3.11.9 |
+| OS | Windows |
+
+# 6. Invariants
+
+1. **Attention backend locked** for a full decode — never flip SDPA↔eager mid-KV cache.
+2. **One WebGL iframe singleton** on `document.body` — Gradio remounts must not spawn new contexts.
+3. **Deterministic forge:** same prompt + same token ids → same forge seed / solid.
+4. **Display budget coherence:** `DISPLAY_POINTS` equals `GPU_SAMPLE_BUDGET`.
+
+# 7. Why It Matters for Field Demos
+
+| Chat shows | MIP shows |
+|------------|-----------|
+| Final tokens | Where attention painted on the parameter cloud |
+| Pass/fail scores | Geometry of a run (forge + compare) |
+| Identity claims | Whether decode focused on expected structure |
+
+MIP is the demo centerpiece alongside Flagship/Growth chat — not a replacement for bakeoffs (Scout/Growth/Flagship **7/8** size-tier; Flagship **10/12** API gauntlet), but the inspectability layer those bakeoffs deserve. Scout remains under-recovered; MIP can still load it for diagnosis without making it a sales hero.
+
+# 8. Limitations
+
+Interactive budgets force sparse sampling — the cloud is not a full weight dump. Attention paint depends on capture hooks and model support. 8 GB VRAM constrains concurrent fleet loads. MIP does not prove answer correctness by itself.
+
+# 9. Conclusion
+
+Model Interior Projection turns decode into a visible, forgeable geometric object. For Arriella’s role-scoped local fleet, that inspectability is a product feature: field and business users can ask not only “what did it say?” but “where did attention land?”
+
+# Acknowledgments
+
+Built with Gradio, WebGL, and PyTorch CUDA stacks on Blackwell consumer hardware.
+
+# References
+
+1. `docs/guides/mip-viewer.md`
+2. Hub MIP note — `docs/papers/hub/arriella-fleet-paper/MIP.md`
+3. Companion factory — `docs/arxiv/01-fleet-factory/`
+4. Companion evaluation — `docs/arxiv/05-evaluation/`
+5. Scripts: `fleet_mip_web.py`, `mip_pipeline.py`, `mip_attention_shape.py`, `mip_viewport_app.py`

papers/README.md

@@ -0,0 +1,56 @@
+# Arriella arXiv manuscript set
+
+**Directory:** `docs/arxiv/`  
+**Purpose:** Publication-ready manuscripts for **arXiv** → **Hugging Face Papers**.  
+**Authors:** Beelzebub4888, Tcoder · **Affiliation:** Infinidev Corp  
+**Contact:** https://formsubmit.co/el/sumuhu · **Hub user:** [UnaverageTech411](https://huggingface.co/UnaverageTech411)
+
+## What this is (and is not)
+
+| Correct | Incorrect |
+|---------|-----------|
+| Submit LaTeX/PDF to **arxiv.org** | Upload papers as Hugging Face **Spaces** |
+| Index announced IDs on **hf.co/papers/<id>** | Fake **model** repos that only contain markdown |
+| Keep Spaces for **runnable demos** only (`Learnit`, etc.) | Treat Spaces as a paper host |
+
+Hugging Face Paper Pages are built on **arXiv**. There is no Hub repo type that replaces arXiv for papers.
+
+## Manuscript index
+
+| ID | Title | Categories | Path |
+|----|-------|------------|------|
+| 01 | Arriella: A Local Fleet Factory for Role-Scoped Small Language Models | cs.CL, cs.LG | [`01-fleet-factory/`](01-fleet-factory/) |
+| 02 | Arriella Grapevine: A Local Omni Multimodal Extension | cs.CV, cs.CL | [`02-grapevine/`](02-grapevine/) |
+| 03 | Honest Capability Ingestion (Eat System) | cs.LG, cs.CL | [`03-eat-system/`](03-eat-system/) |
+| 04 | Multimodal Routing for Text-Only Fleet Tiers | cs.CL, cs.CV | [`04-multimodal-routing/`](04-multimodal-routing/) |
+| 05 | Arriella Fleet Evaluation Report | cs.CL | [`05-evaluation/`](05-evaluation/) |
+| 06 | Fleet Internal Reasoning Wire Format | cs.CL | [`06-reasoning-format/`](06-reasoning-format/) |
+| 07 | Model Interior Projection (MIP) | cs.HC, cs.LG | [`07-mip/`](07-mip/) |
+
+Each folder: `paper.md`, `abstract.txt`, `metadata.yaml` (add `arxiv_id` / `hf_papers` after acceptance).
+
+Sources merged from `docs/papers/`, `docs/guides/`, `docs/benchmarks/`, and `docs/archive/2026-07/`.
+
+## Workflow
+
+Detailed steps: [`SUBMISSION.md`](SUBMISSION.md)
+
+```powershell
+# 1) Build arXiv source zips
+.\.venv\Scripts\python.exe scripts\package_arxiv.py
+
+# 2) Upload each zip at https://arxiv.org/submit  (human step)
+
+# 3) After announcement, index on HF Papers
+.\.venv\Scripts\python.exe scripts\index_hf_papers.py --arxiv-id YYYY.NNNNN --manuscript 01-fleet-factory
+```
+
+Packaged zips land in `docs/arxiv/submission-packages/` (gitignored if large; regenerate anytime).
+
+## Stance rules
+
+- Foundations are clay; Arriella products are factory outputs.
+- Same-size open instruct competition only — no frontier parity claims.
+- Scout under-recovered; Ascension not automatically smarter than Flagship.
+- Grapevine is a multimodal **extension** (no Inkling weights).
+- Prefer measured tables over slogans.

papers/SUBMISSION.md

@@ -0,0 +1,63 @@
+# Submitting Arriella manuscripts to arXiv and Hugging Face Papers
+
+## Why not Spaces or model repos?
+
+- **Spaces** = runnable apps / demos.
+- **Models** = weight checkpoints + model cards.
+- **Papers** on Hugging Face = pages indexed from **arXiv** (`https://huggingface.co/papers/<arxiv-id>`).
+
+There is no supported “upload a PDF as a Paper Page without arXiv” path.
+
+## Step 1 — Package sources
+
+```powershell
+.\.venv\Scripts\python.exe scripts\package_arxiv.py
+```
+
+Creates `docs/arxiv/submission-packages/<manuscript>.zip` containing `paper.tex` (+ `abstract.txt`).
+
+## Step 2 — Submit on arXiv
+
+1. Log in at https://arxiv.org/submit  
+2. For each manuscript, use categories from that folder’s `metadata.yaml`  
+3. Upload the corresponding zip (Process with TeX)  
+4. Paste `abstract.txt` into the abstract field  
+5. Complete license / metadata and submit  
+
+Suggested order: **01-fleet-factory** → **02-grapevine** → **03-eat-system** → **05-evaluation** → shorter companions (04, 06, 07).
+
+## Step 3 — Index on Hugging Face Papers
+
+After the paper is announced (ID like `2607.01234`):
+
+```powershell
+.\.venv\Scripts\python.exe scripts\index_hf_papers.py --arxiv-id 2607.01234 --manuscript 01-fleet-factory
+```
+
+Or open `https://huggingface.co/papers/2607.01234` once to trigger indexing.
+
+Then:
+
+1. Open the paper page → claim authorship if listed  
+2. On future **model** repos, add the arXiv URL in the model card README so HF auto-tags `arxiv:<id>`  
+3. Optionally show the paper on your HF profile (Settings → Papers)
+
+## Step 4 — Link models (when weights are published)
+
+Model cards should cite papers like:
+
+```markdown
+Paper: https://arxiv.org/abs/YYYY.NNNNN
+HF Papers: https://huggingface.co/papers/YYYY.NNNNN
+```
+
+Do **not** create model repos that contain only paper markdown.
+
+## Local SoT
+
+| Path | Role |
+|------|------|
+| `docs/arxiv/*/paper.md` | Editable manuscript source |
+| `docs/arxiv/submission-packages/` | Generated LaTeX zips |
+| `docs/archive/2026-07/` | Historical only |
+| `docs/papers/` | Product whitepaper / drafts (not HF Paper Pages) |

whitepaper.md

@@ -0,0 +1,163 @@
+# Arriella Fleet Whitepaper
+
+**Scout · Growth · Flagship · Ascension — plus Grapevine multimodal & Model Interior Projection**
+
+*Version 4.1 · July 2026*
+
+**arXiv manuscripts (publication SoT):** [`../arxiv/`](../arxiv/) · submit via [`../arxiv/SUBMISSION.md`](../arxiv/SUBMISSION.md) → HF Papers after arXiv ID
+
+---
+
+## Read this first
+
+Arriella is a **local training factory** that turns open instruct foundations into named small LMs you own on disk, plus a **Model Interior Projection (MIP)** viewer that makes inference inspectable in 3D — attention trails on a real parameter cloud, not a black-box chat box.
+
+Organization: **Infinidev Corp** (Infinidev). Lead developers: **Beelzebub4888** and **Tcoder**.
+
+### What problem we solve
+
+1. Useful custom models on a consumer GPU (8 GB Blackwell, Windows)
+2. **Role-scoped tiers** for field and business use — not one generic chatbot
+3. Continuous improvement via **eat** (weight ingest / distill) and **grow** (post-deploy LoRA)
+4. Stability knobs: enterprise base blend, capability/math/thinking anchors, gates before Ollama
+5. A product nobody else ships: **live interior visualization of decode**
+6. A separate **native multimodal** extension when text routing is not enough
+
+### What “Arriella” means
+
+| Sense | Meaning |
+|-------|---------|
+| Project | This repo — scripts, configs, `fleet/`, docs, MIP |
+| Persona | System identity in datasets + chat |
+| Dataset | `arriella_qa` (~90k Pile-derived Alpaca records) |
+| Core products | `arriella-scout`, `arriella-growth`, `arriella-flagship`, `arriella-ascension` |
+| Extension | `arriella-grapevine` (Omni multimodal — not a fifth core text model) |
+
+Arriella is **not** a Meta/Alibaba/DeepSeek foundation model. Foundations are clay; products are what comes out of the factory.
+
+---
+
+## Product thesis — stable models for fields & business
+
+Enterprise and field deployments fail when a single mid-size chat model is asked to be edge-fast, domain-deep, reason-heavy, and multimodal at once. Arriella ships **specialists with shared factory discipline**:
+
+| Need | Route to | Why |
+|------|----------|-----|
+| Low VRAM / high throughput | Scout | Smallest core text tier |
+| Domain QA / continuation | Growth | Trained on full Arriella QA pile |
+| General instruction ops | Flagship | Heretic + premium teacher distill |
+| Explicit step-by-step reasoning | Ascension | DeepSeek-R1 distill native think |
+| Image / audio in the loop | Grapevine | Qwen2.5-Omni-3B student |
+
+Reliability comes from **gates and anchors**, not slogans: capability / math / thinking / identity / fleet-spec datasets, optional enterprise base blend (~8–15% foundation retention), bakeoffs vs same-size stock instruct, and honest modality claims.
+
+---
+
+## Fleet tiers (Jul 2026 stance)
+
+### Core text fleet (exactly four)
+
+| Tier | Size (Ollama) | Foundation | Training path | Product stance |
+|------|---------------|------------|---------------|----------------|
+| **Flagship** | ~1.5B | Qwen2.5-1.5B-Instruct | Heretic → premium QLoRA distill | Primary quality / general ops |
+| **Growth** | ~1.2B | Llama-3.2-1B-Instruct | LLaMA-Factory on `arriella_qa` | Primary mid / domain tier |
+| **Ascension** | ~1.8B | DeepSeek-R1-Distill-Qwen-1.5B | LLaMA-Factory + Flagship catch-up | Reasoning / native think — **not** automatically smarter than Flagship |
+| **Scout** | ~494M | Qwen2.5-0.5B-Instruct | Heretic → QLoRA | Present but **currently underpowered** — edge experiments, not demos |
+
+Canonical numbers and truth rules: [`fleet/CORE_FLEET_SPEC.md`](../../fleet/CORE_FLEET_SPEC.md). Benchmarks: [../benchmarks/README.md](../benchmarks/README.md).
+
+### Multimodal extension (not core text)
+
+| Product | Foundation | Stance |
+|---------|------------|--------|
+| **Grapevine** | Qwen2.5-Omni-3B | Technical preview for vision/audio; modular runtime LoRA + mmproj. Inkling is a capability target only — no Inkling weights. |
+
+Guide: [../guides/grapevine.md](../guides/grapevine.md). HF draft: [grapevine-hf.md](grapevine-hf.md).
+
+---
+
+## Training framework
+
+```
+foundation HF weights
+    ├─ Path A (Scout / Flagship): Heretic abliteration → QLoRA distill → merge
+    ├─ Path B (Growth / Ascension): LLaMA-Factory LoRA → merge
+    └─ Path C (Grapevine): Omni Thinker QLoRA → merge / runtime adapter → GGUF + mmproj
+           ↓
+    fleet/<tier>/model/   (standalone HF)  |  fleet/inkling/ (Grapevine)
+           ↓
+    enterprise base blend (optional) + capability / fleet-spec anchors
+           ↓
+    optional eat (weight ingest / distill) + post-deploy grow
+           ↓
+    GGUF → ollama run arriella-*
+```
+
+**Heretic** abliterates refusal directions (not fine-tuning). Distillation then teaches what to say under QLoRA.
+
+**Pufferfish** in this repo is the **long-run policy / WebUI preset layer** (`fleet_pufferfish.py`, `arriella_integration/`), not the upstream LightLM trainer as the primary engine.
+
+Orchestrator: `scripts/train_fleet.py` + `fleet/fleet_manifest.yaml`.
+
+---
+
+## Eat & grow
+
+After merge, capabilities improve by:
+
+- **Weight ingest** — same-arch tensor merge (linear / TIES / DARE-TIES / …)
+- **Distill grow** — cross-architecture teacher imitation (prefer larger teachers)
+- **Organic grow** — post-deploy LoRA on anchors / datasets
+- **Enterprise align** — base-preservation blend + capability re-eval (`fleet_enterprise_align.py`)
+
+Honest language SoT: [../guides/eat-system.md](../guides/eat-system.md).
+
+---
+
+## MIP — Model Interior Projection
+
+MIP loads a HF/Ollama-backed checkpoint, samples parameter space into ≤120k GPU display points, and during decode:
+
+- Paints prompt-specific attention / residual activity on the cloud
+- Traces a gold polyline (one node per generated token)
+- Forges a deterministic projected **tesseract** from that trail
+- Can **repeat** a prompt N times and **compare** solids (seed, tokens, geometry)
+
+Guide: [../guides/mip-viewer.md](../guides/mip-viewer.md). Entry: `scripts/fleet_mip_web.py`.
+
+This is the differentiator. Chat UIs show tokens. MIP shows **where** attention landed in the model’s own geometry.
+
+---
+
+## Multimodal (honest)
+
+**Core text tiers** do not ingest raw pixels on 8 GB. Images go Gemma3 → categorized text → fleet model. Thinking/vision behaviors are **ingested** via teacher traces.
+
+**Grapevine** is the native Omni path (text + image + audio in Ollama; video validated pre-export). Details: [multimodal.md](multimodal.md), [../guides/grapevine.md](../guides/grapevine.md).
+
+---
+
+## Competitive posture
+
+- Compete with **same-size open instruct** checkpoints via bakeoffs.
+- Do not claim frontier closed-model parity.
+- Prefer measured tables in `docs/benchmarks/` over slogans.
+- Prefer **role fit** (field/business job) over “one model to rule them all.”
+
+---
+
+## Publication (arXiv → Hugging Face Papers)
+
+- Manuscript set: [`../arxiv/`](../arxiv/)
+- Workflow: [`../arxiv/SUBMISSION.md`](../arxiv/SUBMISSION.md)
+- Hub guide: [`../guides/hugging-face-publish.md`](../guides/hugging-face-publish.md)
+
+Do **not** host papers as Spaces or model-only-markdown repos. Spaces are for runnable demos; Paper Pages require arXiv IDs.
+
+---
+
+## Document history
+
+- v2.5 — fleet narrative without MIP as product centerpiece (`docs/archive/2026-07/WHITEPAPER.md`)
+- v3.0 — reorganized docs; MIP as unique product; honest Scout status; three-tier only
+- v4.0 — four-core text fleet + Grapevine extension; field/business thesis; HF paper prep