KnowMe Decision
Self-hosted decision models for KnowMe, as a Rust crate family and MCP sidecar.
know-me-decision is the decision layer shared by KnowMe's products. It replaces hosted "System One" decision APIs such as TypeSafe's Jev with open-weight models running on the user's device or on the clinic's own hardware. Its answers carry calibrated probabilities, and it gives recall guarantees where a missed escalation matters.
Status: pre-implementation
The repository currently holds the design, research and build playbook. Code starts at milestone M0 in the build playbook.
What it answers
- Choice: pick one of up to 255 options.
- Score: an ordinal rubric.
- Noul: the probability that a statement is true.
Every answer comes back as an outcome the host acts on:
| Outcome | Meaning |
|---|---|
Act | One option, calibrated and safe to automate |
Review | A person confirms from a small prediction set |
Escalate | A tripwire or escalation class fired. It is always logged, and no model can override it |
The cascade
Human corrections recorded in the audit log become the labels for the next calibration.
Where it runs
| Host | How it connects |
|---|---|
| The Boss | The bundled knowme-decide sidecar, registered as the built-in MCP server @prometheus/decide |
| Universal Agent Runtime | A linked crate for intent classification, graph routing, model routing and governance context |
| KnowMe (desktop and mobile) | Linked on the device through the gen_ui_decide L2 crate; tripwires and routing work offline |
| Prior Authorization Workbench | A pure aso-decide kernel that proposes met/gap/void and denial paths, and never writes |
| Any MCP client | The knowme-decisions plugin marketplace, plus the ui://decide/review-card MCP App for the human step |
Non-negotiables
- Tripwires run first and can't be overridden.
- No calibration artifact means no automated action.
- Recall floors are set first. The automation rate is measured, never targeted.
- Clinical decisions are proposals. People confirm them through the host's own commands.
- Data never leaves the locality its spec declares.
The full list (I-1…I-10) is in the playbook's invariants.
Read next
- Why not Jev: the question that started the project.
- Open decision models: what to use instead.
- Architecture report: how the layer fits each product.
- Build playbook: how to build it, milestone by milestone.