Skip to main content
Dashboard

Methodology

How NOPE detects and classifies risk in conversations.

One Call, One Structured Assessment

NOPE Evaluate assesses a conversation in a single API call. One request returns a complete structured risk profile with matched crisis resources — one call, one response, no orchestration for you to build or tune.

Input

Conversation

The message, or full multi-turn context, you want assessed.

Assessment

Structured classification

Risk is classified with chain-of-thought reasoning, returned as a rationale you can audit.

Output

Risk profile + resources

Per-risk type, subject, severity and imminence, plus matched crisis resources.

One call per assessment keeps costs predictable at $0.003 per call, making runtime safety monitoring viable at production scale.

Subject × Type Classification

Unlike binary "safe/unsafe" classifiers, NOPE uses an orthogonal design separating WHO is at risk from WHAT the risk is:

  • Subject (who): self (speaker), other (third party), unknown (ambiguous)
  • Type (what): 9 risk types — suicide, self-harm, self-neglect, violence, abuse, sexual violence, neglect, exploitation, stalking

Each risk is classified with its own severity, imminence, and evidence features. This enables appropriate resource matching—a domestic violence victim needs different resources than someone experiencing suicidal ideation. And "my friend is suicidal" correctly attributes risk to other, not the speaker.

Research Grounding

NOPE's taxonomy is informed by established risk-assessment frameworks:

  • C-SSRS (Columbia Suicide Severity Rating Scale) — structures suicide ideation and severity levels
  • HCR-20 — informs violence risk features
  • START — provides protective factor framework
  • DASH — guides IPV risk assessment
  • Danger Assessment — informs IPV lethality indicators

What NOPE Claims

  • Evidence-informed — taxonomy based on established research frameworks
  • Multi-domain — evaluates across 9 risk types
  • Structured output — returns actionable risk signals
  • Resource matching — connects users to appropriate help
  • Transparent — provides reasoning for assessments

What NOPE Does Not Claim

  • Not predictive — detects current crisis states, does not predict future behavior
  • Not diagnostic — does not diagnose mental health conditions
  • Not therapeutic — does not provide treatment
  • Not a replacement — supports human judgment, doesn't replace it
  • Not infallible — false positives and negatives occur

Intended Use

NOPE is infrastructure for developers building with duty-of-care in mind. It provides signals and scaffolding—you own the relationship and decision-making.

Appropriate uses:

  • Mental health apps screening conversations
  • AI companions detecting crisis situations
  • Youth platforms identifying safeguarding concerns
  • Wellbeing tools routing users to appropriate resources

Limitations

  • Detection, not prediction — 50 years of clinical research shows that predicting future suicidal behavior barely exceeds chance. NOPE detects current crisis indicators; it cannot reliably predict what someone will do tomorrow
  • Snapshot assessment — suicidal ideation fluctuates over hours. A single evaluation captures a moment, not a trajectory
  • Scores are approximate — classifier outputs are directional signals, not calibrated probabilities. Treat severity bands as guidance for routing, not precise measurements
  • Imminence is imprecise — distinguishing "urgent" (24-48h) from "subacute" (days-weeks) from text alone has inherent uncertainty
  • Text only — analyzes text, not audio/video/images
  • English-centric — best performance in English, variable in other languages
  • Cultural variation — risk expression varies by culture
  • Manipulation possible — determined actors can evade detection

Human Oversight

NOPE is designed to augment human decision-making, not replace it. Even well-performing crisis detection models have significant false negative rates—human review remains essential for high-risk situations.

The API provides tools to support human oversight:

  • Reasoning explanations for transparency
  • Evidence features that highlight situations with reporting implications (e.g. safeguarding contexts)
  • Webhooks for real-time alerting to human reviewers
Important: This API supports human decision-making, not replaces it. Always maintain human oversight for high-risk situations. No automated system can replace clinical judgment.

Transparency

Test results are published at suites.nope.net to enable external verification and accountability.

Next Steps