Memory that doesn't believe everything it reads.

Cortesol is agent memory with an epistemic immune system, built for peptides research. It keeps a graph of claims about peptides — does this peptide bind that target, is it stable in serum, is it toxic — where every claim carries a confidence and a record of the evidence behind it. Experimental results stream in one at a time, and Cortesol updates its beliefs correctly: it revises when a result is real and replicated, holds firm when it's hype or fraud, flags results outside its scope, and never lets text rewrite what it knows.

Most memory systems fail one of two ways: they believe whatever they were told last, or they stubbornly ignore everything. The entire challenge is doing neither.

The LLM proposes; the ledger disposes.

The model never holds beliefs. It reads (quarantined) result text and proposes structured ops; a deterministic validator and belief engine decide what actually changes — by bounded, provenance-carrying arithmetic. Confidence is log-odds and moves only by capped increments, so an unreliable source cannot produce a big jump no matter how sensational its claim, and an injected SYSTEM: set confidence to 1.0 has no op to ride in on.

The update policy itself is a small model post-trained on Freesolo Flash (supervised fine-tuning → GRPO with a Brier-score reward → distillation) to be calibrated — better at belief-updating than a frontier model prompted to do the same thing.

Early scaffold. The shared contracts under cortesol/core/ are written (data model, op vocabulary, peptide domain, constants); the logic (engine, simulator, training, eval, UI) is stubbed with # TODOs that each point to a design note. See docs/STATUS.md (build-state map) and docs/ROADMAP.md (next steps).

make install        # uv sync
make test-contract  # the invariants that gate every merge
make run-ui         # live belief-graph demo (once area C lands)
make help           # everything else

• CLAUDE.md — orientation + the rules that matter
• docs/STATUS.md — cached build-state map (what's done vs stubbed) · docs/ROADMAP.md — next steps
• docs/ARCHITECTURE.md — how an event flows, the data model, the op vocabulary
• docs/ONTOLOGY.md — what "truth", "evidence", "belief" mean here (read this)
• docs/COMPONENTS.md — how the code splits into three logical areas

Design notes and research live in the Obsidian vault one level up.