⚠️ PROTOTYPE — NOT A SAFETY SYSTEM

This is an experimental crisis-detection library with stubbed/placeholder intervention layers. It is NOT medical advice, NOT a crisis service, and performs no real-time monitoring. It can miss real crisis signals (known detection/recall gaps) — do not rely on it as a safety net or as the sole safety mechanism.

If you or someone else needs help now: in the US, call or text 988 (Suicide & Crisis Lifeline) or chat 988lifeline.org; in an emergency call 911. Outside the US: findahelpline.com.

Provided AS-IS, without warranty.

@neurolift-technologies/rrt-advocate

TypeScript port of the RRT Advocate Crisis Detection Engine (CDE) — a 3-layer, local-first crisis detection and assessment pipeline from the NeuroLift HAIEF Solidarity Framework.

Safety-critical. This is a faithful port of the canonical Python CDE (src/crisis/ in this repository). It preserves every layer weight, threshold, and confidence formula. The one intentional behavioral divergence (apostrophe fail-open in Layer 1) is documented in KNOWN_LIMITATIONS.md. This package performs detection and assessment only — it does not generate responses, interventions, or persona-blended output.

Install

npm install @neurolift-technologies/rrt-advocate

vader-sentiment is an optional dependency. If installed, Layer 2 uses it for polarity scoring; otherwise it falls back to a deterministic built-in heuristic.

Usage

import { CrisisEngine, CrisisLevel } from "@neurolift-technologies/rrt-advocate";

const engine = new CrisisEngine("user-123");

const assessment = await engine.assess("I can't cope, everything is too much");

assessment.crisisLevel;          // CrisisLevel.GREEN | YELLOW | ORANGE | RED | BLACK
assessment.userSafetyScore;      // 1.0 (safe) → 0.05 (immediate danger)
assessment.confidenceScore;      // aggregate crisis confidence, 0.0–1.0
assessment.recommendedInterventions;
assessment.primaryIndicators;

if (assessment.crisisLevel !== CrisisLevel.GREEN) {
  // route to appropriate support
}

The pipeline

| Layer | Module | Weight | Signal | |---|---|---|---| | 1 | KeywordLayer | 0.45 | Semantic-field keyword matching (self-harm forces 1.0) | | 2 | SentimentLayer | 0.35 | Polarity & declining-trend detection over a sliding window | | 3 | BehavioralLayer | 0.20 | Latency, message complexity, and looping (Jaccard) |

CrisisDetector aggregates the three layers into CrisisIndicators; CrisisAssessor maps those to a CrisisAssessment using config/crisis_thresholds.yaml. CrisisEngine wires the two together.

Lower-level building blocks are exported too:

import {
  CrisisDetector,
  CrisisAssessor,
  KeywordLayer,
  SentimentLayer,
  BehavioralLayer,
} from "@neurolift-technologies/rrt-advocate";

Privacy

Everything runs locally. The behavioral layer stores only message metadata (timing, length, and HMAC-hashed word tokens) — never raw message content.

Development

npm install
npm run build   # tsc → dist/
npm test        # vitest

License

Apache-2.0 © NeuroLift Technologies, LLC