kaOS IntelligencekaOS LIBRARY · 2026
ARCHITECTUREABOUTVISIT THE SHOWROOMTALK TO US
kaOS IntelligenceDOC 004 · BOOK EVOLVE · REV 2026.07
¶ 1 · THE INSTRUMENT

A developmental editor that argues with itself until the prose is yours.

Book Evolve is a command-line tool that turns raw dictated transcripts into finished book chapters. Three models draft in parallel. A judge scores them on six axes. The winner goes back for another round, and another, until the work clears the bar — in your voice, not the model's.

ONE COMMAND · NO CLONE · NO BUILD

COUNCIL · 3 MODELS
Three different models draft the same chapter, blind to each other. Disagreement is the point.
JUDGE · 6 AXES
Content fidelity, style match, literary quality, structure, readability, novelty. Verbatim copies score zero.
VOICE · YOURS
Seed it with your own reference prose. The judge learns your voice and refuses to let the model talk over it.
MACOS / LINUXNODE 18+PYTHON 3.10+BRING YOUR OWN MODELSOPENROUTER · OLLAMA · LITELLMMIT
I. How it works

Four stages. One chapter at a time.

Every chapter walks the same path. The council never sees the judge; the judge never sees the council's identities. Only the prose crosses the line.

01
Seed
Drop in a raw dictation, a reference chapter, or a style sample. The cognition store holds it whole.
02
Provision
Three models draft the chapter in parallel, each from the same seed, each blind to the others.
03
Evolve
A judge scores every draft on six axes, picks the strongest, and sends it back for another round.
04
Promote
When a draft clears the target score — or patience runs out — the best version is promoted to output.
II. The council and the judge

One model's first draft is never the answer.

The council drafts in parallel. The judge reads every draft cold — no model names, no ranking bias — and scores it on six axes. The strongest goes back for another evolution round.

COUNCIL · DRAFTS IN PARALLEL
Model 1
COUNCIL_MODEL_1
Model 2
COUNCIL_MODEL_2
Model 3
COUNCIL_MODEL_3
The Judge
JUDGE_MODEL · SCORES ON SIX AXES
Content fidelity
Did it say what the dictation said?
Style match
Does it sound like the author, not the model?
Literary quality
Is the prose itself worth reading?
Structural coherence
Does the chapter hold together?
Readability
Does a sentence earn its length?
Evolutionary novelty
Is this draft meaningfully different from its siblings?
↓ PROMOTES THE STRONGEST · SENDS IT BACK FOR ANOTHER ROUND
III. Five minutes to your first chapter

From npx to a promoted chapter.

No clone, no build, no account. Bring an OpenAI-compatible endpoint and a dictation; Book Evolve does the rest.

01
Install
No clone, no build. Run it once and it scaffolds a project in your current directory.
npx @kaos-intelligence/book-evolve
02
Scaffold
Name the book and the author. This writes config.yaml, .env.ai.example, and the project skeleton.
book-evolve scaffold --title "Your Book" --author "Your Name"
03
Point it at a model
Edit .env.ai with any OpenAI-compatible endpoint. OpenRouter is the fastest way to start.
LITELLM_BASE_URL=https://openrouter.ai/api/v1
LITELLM_API_KEY=sk-or-v1-...
COUNCIL_MODEL_1=anthropic/claude-3.5-sonnet
COUNCIL_MODEL_2=google/gemini-flash-1.5
COUNCIL_MODEL_3=openai/gpt-4o-mini
JUDGE_MODEL=anthropic/claude-3.5-sonnet
FAST_MODEL=openai/gpt-4o-mini
04
Evolve
One command. The council drafts, the judge scores, the chapter promotes when it clears the bar.
book-evolve evolve --chapter 1

STUCK? RUN book-evolve doctor — IT TELLS YOU EXACTLY WHAT'S MISSING.

IV. Bring your own models

Your keys. Your endpoints. Your cost.

Book Evolve never calls a model it wasn't pointed at. Drop one of these into .env.ai and you're running. The defaults in the scaffold are placeholders — override them.

THE FASTEST START
OpenRouter
One key, every major model. Pay per token. No local setup beyond Node and Python.
LITELLM_BASE_URL=https://openrouter.ai/api/v1
LITELLM_API_KEY=sk-or-v1-...
COUNCIL_MODEL_1=anthropic/claude-3.5-sonnet
COUNCIL_MODEL_2=google/gemini-flash-1.5
COUNCIL_MODEL_3=openai/gpt-4o-mini
JUDGE_MODEL=anthropic/claude-3.5-sonnet
FAST_MODEL=openai/gpt-4o-mini
FULLY LOCAL, FULLY PRIVATE
Ollama
Nothing leaves your machine. Pull the models, point the CLI at the local server, evolve offline.
LITELLM_BASE_URL=http://127.0.0.1:11434/v1
LITELLM_API_KEY=ollama
COUNCIL_MODEL_1=qwen3:32b
COUNCIL_MODEL_2=llama3.3:70b
COUNCIL_MODEL_3=gemma3:27b
JUDGE_MODEL=qwen3:32b
FAST_MODEL=qwen3:8b
OLLAMA_EMBED_MODEL=bge-m3:latest
YOUR OWN ROUTER
LiteLLM proxy
Already running LiteLLM? Point at it. Master key resolves from ~/.litellm-master-key if set.
LITELLM_BASE_URL=http://127.0.0.1:4000/v1
# LITELLM_API_KEY read from ~/.litellm-master-key
COUNCIL_MODEL_1=your-route-alias-1
COUNCIL_MODEL_2=your-route-alias-2
COUNCIL_MODEL_3=your-route-alias-3
JUDGE_MODEL=your-route-alias-1
FAST_MODEL=your-fast-route

No embeddings server running? Set ASI_EVOLVE_HASH_EMBEDDINGS=1024 to evolve with degraded hash retrieval — slower retrieval, same prose. The full environment reference is in the README.

V. Observability

Every evolution leaves a paper trail.

Each run appends structured events to activity_stream.jsonl — provision attempts, per-axis judge scores, promotion decisions. Best-effort by design: a logging failure can never take down a run.

{"ts":"2026-07-19T18:42:11Z","event":"judge.score","chapter":1,"step":3,"model":"claude-3.5-sonnet","axes":{"content_fidelity":0.91,"style_match":0.88,"literary_quality":0.86,"structural_coherence":0.9,"readability":0.92,"novelty":0.74},"promoted":false}
{"ts":"2026-07-19T18:43:54Z","event":"chapter.promoted","chapter":1,"step":5,"final_score":0.94}

Bring us the manuscript you've been dictating for years.

© 2026 KAOS INTELLIGENCE, INC.HOMEARCHITECTUREBOOK EVOLVE ON GITHUBEND OF DOCUMENT
¶ THE LIBRARY, OPEN ALL NIGHT

Send us a document. Any document.

We'll return it thinking.

Talk to usRead today's Spark
© 2026 KAOS INTELLIGENCE, INC. · KNOWLEDGE-AWARE OPERATING SYSTEMARCHITECTUREABOUTAI AGENT? READ /LLMS.TXT · /LLMS.MD · /MCPEND OF DOCUMENT · REV 2026.07