Current seven-reflex APIs, confidence semantics, lifecycle boundaries and defaults are documented in the integration guide. Examples below describe the original minimal integration.

@brainstem/core

The reflex engine underneath brainstem. Most consumers should use @brainstem/reflexes instead — this package is the low-level foundation: the SystemOne judge interface, ReflexEngine, policy, the journal format, and the bitmap/section utilities that back Focus.

What's here

• SystemOne — the judge interface every reflex asks questions through: { name, ask(state, questions, options) => Promise<AskResult> }. Implement this against any model.
• ReflexEngine — orchestrates Gate, Sanitize, Verify, Pulse, Steer, and Focus over a SystemOne, a Policy, and a Journal.
• policyForTrust(trust) — safety thresholds that scale with a single trust dial (0–1), rather than exposing a pile of independent knobs. Thresholds near the safety-critical end are trust-invariant by design. See src/policy.ts.
• openJournal / loadJournal — a schema-versioned NDJSON decision log, replayable offline via @brainstem/cli's replay tooling (not published here).
• splitIntoSections / CapabilityBitmap — structural output splitting and the LSB0 bitmap format Focus uses to mark which sections were selected.

Stability

This package's API surface is broader and less stable than @brainstem/reflexes. If you're building on brainstem, start with @brainstem/reflexes and only reach into core directly if you need something the higher-level API doesn't expose.

License

MIT — see the repo root LICENSE.