Decision kernel

The decision kernel is Rna’s only path to JEV. The code lives in four places:

Modes and backendsCopy Anchor Link

• Modes: off, shadow (record only, no behavior change), active.
• Backends: typesafe (JEV) or llm (a language model acting as judge).
• Each result names its source: judge for a judgment, fallback for the rule used when no judge is available, and rule for a deterministic rule of the host.
• In the Rna app all decision points are built in and active by default, with only a few off or shadow; mode and backend can only be overridden for diagnosis with the environment variables RNA_JEV_MODE, RNA_JEV_MODES and RNA_JEV_BACKEND. The backend is chosen automatically: the JEV key first, then the fast model, and with the fast model (llm) every threshold is raised by 0.1.

ThresholdsCopy Anchor Link

Thresholds are tiered by consequence; the constant is DECISION_THRESHOLDS in decision-templates.mjs. A yes/no question’s confidence is max(p, 1 - p):

The constants also hold a floor of 0.5, the fast-model offset llmOffset of 0.1, and numeric thresholds set per decision point (some points use two, {act, ask}). Look at the real distribution on a labelled set before choosing a tier for each question. Never copy one from another question or type.

Adding a decision pointCopy Anchor Link

1. Write the question template and its combining function in decision-templates.mjs. Phrase questions as literal, observable situations, one judgment each.
2. Register the template with every question ID in TEMPLATE_SPECS. Missing entries make the kernel reject the whole plan as invalid_assessment_plan; packages/sdk/test/decision-templates-valid.test.mjs guards this.
3. Register the point in the host's decision-service.mjs (title, risk tier, effect) so it appears in the settings Judgment records.
4. Build the state in decision-state.mjs: only the text the judgment needs, with lengths, counts and dates computed in code and bucketed.
5. Run a labelled set through the probe, record the distribution next to the study, then choose thresholds.

ProbeCopy Anchor Link

POST /api/v2/decision-kernel/probe
Content-Type: application/json

{ "state": "…", "questions": { "…": { "type": "noul", "instructions": "…", "criteria": "…" } } }

The probe sends one raw call with the app’s configured key, bypassing thresholds, cache and ledger. The reusable harness is prototypes/control-center/benchmark/jev-capability-probe.mjs; set RNA_STATE to the main checkout’s .state/state.json to read real rule bodies.