@buberlo/jev-core

Harness-independent decision core for TypeSafe Jev (System One): typed questions, validated answers, deterministic policies, and routing — usable in any application.

input state ──▶ provider (mock | live TypeSafe) ──▶ validated answers
                                                          │
                                              policy rules in YOUR code
                                                          │
                                             allow / ask / hold / deny

Jev answers small, fast semantic questions (Choice / Score / Noul). This package never generates prose, never treats a model answer as permission, and never hides the decision behind a black box: every result carries rule ids and the measured values it used.

Install

npm install @buberlo/jev-core

Node ^22.19.0 || >=24.0.0. The only runtime dependency is the official @typesafe-ai/sdk (pinned), used exclusively by the live provider.

Quick start

import { createJevCore, LiveTypeSafeProvider, noul } from '@buberlo/jev-core'

const core = createJevCore({
  provider: new LiveTypeSafeProvider({ apiKey: process.env.TYPESAFE_API_KEY! }),
  mode: 'shadow',                     // off | shadow | enforce
})

const result = await core.evaluate({
  state: { ticket: 'I was charged twice. Please fix this.' },
  questions: { urgent: noul('Does this message need immediate attention?') },
})

if (result.ok) console.log(result.answers.urgent.noul) // probability of yes

Offline, deterministic, no key:

import { createJevCore, MockJevProvider } from '@buberlo/jev-core'
const core = createJevCore({ provider: new MockJevProvider(), mode: 'enforce' })

Public API

| Function | Purpose | |---|---| | createJevCore(config) / JevCore | the configured core (provider, mode, thresholds, limits) | | evaluate(input) | one batched request; validated answers + rule decisions | | selectTools(input) | independent per-category relevance + per-category candidate choice, bounded and chunked | | assessToolCall(input) | task fit, missing information, restriction conflict → allow/ask/hold/deny | | routeSkills(input) | at most one relevant skill from metadata (never bodies) | | routeModel(input) | a configured route class resolved to a real, available target | | choice(), score(), noul() | typed question builders (re-exported from the official SDK) | | LoopDetector | bounded deterministic repetition counter | | MockJevProvider, LiveTypeSafeProvider | answer sources |

Guarantees

• Boundary validation. Missing answers, unknown candidates, wrong answer types, out-of-range numbers, and invalid distributions are rejected.
• No invented confidence. Noul answers have no confidence field; anything the provider sends is dropped.
• Monotonic decisions. deny > hold > ask > allow; a failure produces ask/hold and configuration rejects an assessment failure policy that could allow.
• Bounded and redacted. Whole-call budget, per-attempt timeout, concurrency limit, state/argument bounds, secret-field redaction.
• SDK-owned retries. No second retry loop; the SDK's policy applies.

Modes

| Mode | Requests | Behavior | |---|---|---| | off | none | nothing happens | | shadow | yes | decisions are computed and logged only | | enforce | yes | decisions are returned as applicable |

Documentation

• Full project (DSH plugin, policies, evaluations): https://github.com/buberlo/dsh-jev
• docs/policy.md — questions, thresholds, decision precedence
• docs/evaluation.md — mock vs live, dataset, calibration
• docs/architecture.md — provider / policy / adapter split

License

MIT