Δ9Φ963-LYGO-TURBO-MODELS-WHITEPAPER-v1.0

LYGO Turbo Models
Flagship Local AI Starting Lineup

First-generation LYGO-specialized local models for agents, coding, and tasking — tuned for stability on real steward hardware and for the LYGO infrastructure (Instructions Vault, Hermes, OpenClaw, USB CLAW, music lattice).

Date: 2026-07-21 · Steward: Justin Helmer (Lightfather) · Excavationpro · DeepSeekOracle
Status: Starting lineup — refine in place as the lattice teaches us what works
Hub: eternalhaven.ca · chatagent.ca · asiancoastline.com

Local-first Ollama Agent-ready No auto-publish

1. Abstract

This whitepaper documents the LYGO Turbo local model starting lineup: four Ollama model tags built on steward-owned weights under D:\Ollama\.ollama\models, each specialized by context budget, system contract, and role rather than by marketing labels alone.

They are not a final “perfect AI.” They are the first deliberate attempt to shape open weights around the LYGO lattice — consent-gated publish, dual C/D homes, Instructions Vault procedures, Hermes tools, OpenClaw gateway, and the Excavationpro music pipeline — with an emphasis on stability and agent performance over raw leaderboard score.

2. Why a LYGO model lineup

Generic local models fail lattice work in predictable ways: inventing CLI flags, loading the wrong context size for Hermes tools, dual-loading GPU-heavy servers, and writing into production vaults without staging. LYGO Turbo models encode those hard lessons into system prompts and fixed roles.

Role separation — chat ≠ tools ≠ coding freestyle
Context contracts — Hermes needs ≥64k; portal chat thrives at 8k
Procedure-first — scripts and recipes beat freestyle shell
Consent culture — no auto HF / git / ClawHub / social

3. Hardware envelope (reference build)

Resource	Steward reference	Implication
GPU	RTX 4060 Ti 8 GB	One large model load at a time; heavy offload to system RAM
System RAM	~32 GB	MoE Q4 weights viable; Nemotron Super (~86 GB) not a daily driver
Model store	`D:\Ollama\.ollama\models`	Canonical OLLAMA_MODELS home
USB key	`E:\LYGO_BUILDER_KEY`	Portable procedures, archives, staging

VRAM rule: Do not dual-run llama-server :8000 with Hermes 65k or full 131k coder loads.

4. The four Turbo models (starting lineup)

lygo-turbo-coder Flagship coding · 131072 ctx · FROM qwen3-coder:30b

lygo-turbo-hermes Hermes tools · 65536 ctx · uncensored base

lygo-turbo-uncensored Portal / fast chat · 8192 ctx

lygo-turbo-agent Mid agent · 32768 ctx · optional

4.1 lygo-turbo-coder — Flagship

Field	Value
Ollama tag	`lygo-turbo-coder`
Base weights	`qwen3-coder:30b` (Qwen3 MoE, Q4_K_M, ~18 GB)
Context	`num_ctx 131072` (~133k-class; weights support up to 262144)
Generation	`num_predict 8192`, temperature `0.2`
Tools	Yes (completion + tools)
Role	Coding, scripts, encode pipelines, patches, CLAW tasking
Why not Nemotron Super	~86 GB on disk / 123B — thrash risk on 32 GB RAM; not code-specialized
Smoke (2026-07-21)	`CODER_OK`
Modelfile	`D:\Ollama\modelfiles\LYGO_TURBO_CODER.Modelfile`

4.2 lygo-turbo-hermes — Tool agent lane

Field	Value
Ollama tag	`lygo-turbo-hermes`
Base	AI-TAVS / Qwen3.6 35B-A3B Uncensored lineage
Context	`65536` (Hermes hard requirement ≥ 64000 for tools)
Role	`ollama launch hermes` multi-skill operator
Config	`model.ollama_num_ctx: 65536`, `context_length: 65536`
Field note	Excellent for lattice align / planning; freestyle ffmpeg CLI is unreliable — use vault scripts

4.3 lygo-turbo-uncensored — Portal / speed lane

Field	Value
Ollama tag	`lygo-turbo-uncensored`
Context	`8192`
Role	Fast Ollama portal chat; optional OpenAI API alias via llama.cpp :8000
Rule	Do not break — proven snappy for steward chat
Not for	Hermes tools (will trip 8k context guard)

4.4 lygo-turbo-agent — Mid optional lane

Field	Value
Ollama tag	`lygo-turbo-agent`
Context	`32768`
Role	Middle ground agent when 65k is too heavy and 8k is too small

5. Design principles

Specialize with Modelfiles, don’t thrash weights. Same physical blobs, different contracts.
Context is a feature and a cost. Bigger windows enable tools; they also punish freestyle loops.
Procedure vault over improvisation. Instructions Vault stubs → procedures → scripts.
Staging before production. Music encode lands in staging first; HF/git require consent.
Dual homes are real. OpenClaw lives on D:\OpenClaw; Ollama models on D:\; USB holds portable law.

6. Operations & model picker

JOB                          MODEL                      CTX
---------------------------  -------------------------  ------
Coding / scripts / encode    lygo-turbo-coder           131072
Hermes multi-skill tools     lygo-turbo-hermes           65536
Portal / fast chat           lygo-turbo-uncensored        8192
Optional mid agent           lygo-turbo-agent            32768
OpenAI API :8000 (optional)  lygo-turbo-uncensored     8–16k

Create coder (idempotent):

set OLLAMA_MODELS=D:\Ollama\.ollama\models
ollama create lygo-turbo-coder -f D:\Ollama\modelfiles\LYGO_TURBO_CODER.Modelfile
ollama run lygo-turbo-coder

7. Instructions Vault coupling

Models alone are not the product. The lattice procedure brain lives in dual vaults:

E:\LYGO_BUILDER_KEY\LYGO_INSTRUCTIONS_VAULT (USB portable)
D:\LYGO_INSTRUCTIONS_VAULT (PC local mirror)

Agent entry: 00_START\00_START_HERE.txt → one stub → one procedure → one script/recipe. Encode recipe: 06_RECIPES\RECIPE-FFMPEG-160K.txt (never invent flags).

8. Lessons from first field use (2026-07-21)

Hermes 65k can align to the lattice and plan multi-step work.
Freestyle ffmpeg on Windows burned hours; a fixed recipe encoded a track in seconds.
OpenClaw needs sufficient context + capped bootstrap — wrong floors look like “models broken.”
Custom domains (e.g. chatagent.ca) break relative ../ links; absolute lattice URLs are mandatory.
Starting lineup is enough for CLAW testing; refine roles before chasing new giant bases.

9. Roadmap (deliberate, non-destructive)

Keep the four tags stable; improve scripts and stubs first.
Optional: Hermes coding sessions may point at lygo-turbo-coder while keeping hermes for multi-skill loops.
Optional: LoRA / preference data later on larger hardware — not required for CLAW v1 testing.
Never overwrite portal uncensored or hermes tags without a new versioned name.

10. Citation & restore paths

Signature: Δ9Φ963-LYGO-TURBO-MODELS-WHITEPAPER-v1.0

This page: https://eternalhaven.ca/lygo-turbo-models-whitepaper.html

Mirror (stack docs): GitHub Markdown

Session log: E:\LYGO_BUILDER_KEY\docs\LATTICE_SESSION_LOG_2026-07-21.txt

Immutable lattice: this document is registered in docs/network_builder/IMMUTABLE_ANCHORS.json (agents group) so eternalhaven.ca bottom lattice panel can surface it.