A local-first, self-improving knowledge system that turns school lessons into a queryable Source of Truth — engineered as notes for an open-book exam where LLMs are disallowed. The AI does the capture weeks early; the cold notes are the only aid you carry in.
A study tool that hallucinates is worse than none — so my-AI-stro makes verification a system property: every lesson is captured into a structured SOT, and a deterministic (non-LLM) judge scores grounding at every write. The crowning jewel of the MoreSalamander methodology — the reference implementation a stranger can read to understand what the thesis means in practice.
The lesson, the SOT schema, and the grounding criteria the judge will apply — written down first. Trust isolation is part of the doctrine: the model that owns the SOT never handles ungrounded chat.
A local model (llama3:8b) summarizes the lesson into structured SOT entries; retrieval pulls related prior entries so the new entry fits what's already known.
A deterministic, formula-based judge scores grounding against the source and rejects ungrounded entries. A background audit loop keeps rotating canonical entries toward more-grounded versions over time.
my-AI-stro runs a full curation chain (Ingest → SOT → Chat → Notebook → Classroom), each boundary Python-verified, with three coexisting pipelines on one shared NDJSON event vocabulary and a gradebook that extends verification into learning analytics. Everything local; the notes are yours and stay on your machine.