An unpublished manuscript stays on your computer
This hosted Preflight runs on a free AI tier whose terms let the provider use what you send it, so it takes public-domain books and the demo series only. Preflight is open source: the same pages, the same series bible and the same report run on your own computer with a local model, and your text never leaves it.
Run Preflight on your own computer
- Install Python 3.11 or newer and Ollama, then fetch a model:
ollama pull qwen3:4b(2.5 GB; a laptop without a graphics card runs it). - Start Ollama with a large enough window:
OLLAMA_CONTEXT_LENGTH=16384 ollama serve - Install Preflight:
pip install git+https://github.com/nolankido/canonKeeper - Start it:
canonkeeper serve --provider openweights, openhttp://127.0.0.1:8787, and upload your books as you would here. Nothing is charged: payments are simulated on your own copy.
Any OpenAI-compatible server works in place of Ollama (LM Studio, vLLM, a GPU box on your network): set CANONKEEPER_BASE_URL, CANONKEEPER_EXTRACT_MODEL and CANONKEEPER_ADJUDICATE_MODEL.
What a small local model gets right, measured
On the three-book demo series with five planted errors, a 4-billion-parameter model on a laptop CPU built a bible in which every fact is tied to a verbatim sentence, and three of the five errors reached the list of conflicts. It is not a good judge: it called all three explainable. Preflight does not let a model clear a conflict without quoting the passage that explains it, so those three came back as Needs your judgement for you to read, not dismissed. So a small local model gives you a private, sourced bible and a list of conflicts to read yourself, not verdicts to trust (the full measurement).