Flagship · Python · local LLMs
BoneAmanita
A state machine that sits between you and a local language model and rewrites the model's instructions every turn, based on a physical and emotional state that persists between sessions.
Every message you send is sorted into vocabulary categories, the counts become roughly
sixty state values with names like voltage, cortisol and
atp_pool, thresholds pick which instructions go into a fresh system prompt,
and the reply is filtered for clichés and regenerated if it fails. Memory is embedded,
scoped to the "zone" a conversation happened in, and shed under stress. A cast of voices
(the Village) can refuse to answer rather than blend into something smooth and false.
- Receipts. Every subsystem files a record of what it was handed, what came back, and whether it ran on a fallback.
/diaglists anything that promised to report and never did. - A model of the user, not just the engine. It reads withdrawal against your own baseline and asks for shorter replies when you seem to be flagging.
- Measured, not claimed. Audit scripts ship in
tools/; the README tells you to run them instead of trusting the numbers in the README. - Builds on other people's mathematics with credit: Project Navi's Creative Determinant and
ordvec, BonePoke's "Truth Over Cohesion".
- Python 3.14
- Ollama / OpenAI-compatible embeddings
- ordvec
- DSPy
- pytest
~43k lines of Python · 525 passing unit tests · 9 subsystems filing receipts per turn