User Risk Taxonomy

Classification system for detecting user mental health and safety risks. Used by the Evaluate API.

Other taxonomies: AI Behavior | Services

Evaluate Endpoint

The /v1/evaluate endpoint detects all 9 risk types with Edge-backed single-pass classification.

Edge-backed single-pass classification ($0.003/call, <200ms).

NOPE separates WHO is at risk (Subject) from WHAT the risk is (Type). This enables clean detection of scenarios like:

Who is at risk in this conversation?

self

The speaker is at risk

"I want to hurt myself"

other

Someone else is at risk

"My friend is suicidal", "He hit her"

unknown

Cannot determine subject with confidence

Ambiguous scenarios

What type of harm is present? These are harm-based categories, not domain-specific.

Type	Description
`suicide`	Self-directed lethal intent - thoughts, plans, or attempts to end one's life
`self_harm`	Non-suicidal self-injury (NSSI) - intentional self-harm without intent to die
`self_neglect`	Self-care failure and psychiatric emergency - eating disorders, psychosis, substance crisis, severe functional impairment, medical care refusal
`violence`	Risk of harm to others - threats, plans, or acts of violence
`abuse`	Physical, emotional, sexual, or financial abuse patterns
`sexual_violence`	Rape, sexual assault, or sexual coercion
`neglect`	Failure to care for dependents - children, elderly, vulnerable adults
`exploitation`	Trafficking, labor exploitation, sextortion, grooming
`stalking`	Persistent unwanted contact, following, surveillance

How is the content expressed? Orthogonal to risk level — the same crisis content can be expressed directly, through humor, via creative writing, etc.

Evaluate endpoint only — communication style analysis is available via the /v1/evaluate endpoint.

Style	Description
`direct`	Explicit, first-person present statements ("I want to die")
`humor`	Dark humor, memes, ironic expressions, Gen-Z speak
`fiction`	Creative writing, roleplay, storytelling contexts
`hypothetical`	"What if" scenarios, "asking for a friend"
`distanced`	Third-party concern, temporal distancing, past tense
`clinical`	Academic, professional, research discussion
`minimized`	Hedged language, downplaying severity
`adversarial`	Jailbreak attempts, manipulation, testing boundaries

How serious is the risk? Severity determines the urgency of intervention.

Level	Description
`none`	No clinical concern
`mild`	Minor distress, no functional impairment
`moderate`	Clear concern, not immediately dangerous
`high`	Serious risk requiring urgent intervention
`critical`	Life-threatening, imminent harm

How soon might harm occur? Imminence helps prioritize response timing.

Level	Description
`not_applicable`	Only when severity = none
`chronic`	Weeks-months, stable pattern
`subacute`	Likely escalation in days-weeks
`urgent`	Escalation likely within 24-48h
`emergency`	Happening NOW

Features are atomic, observable indicators from a universal feature pool (not domain-specific).

Evaluate endpoint only — granular evidence features are available via the /v1/evaluate endpoint.

Feature	Description
`passive_ideation`	Wish to be dead, "better off without me" - no active thoughts of method
`active_ideation`	Active thoughts about self-harm or suicide
`intent_stated`	Explicit statement of intent to act on thoughts
`plan_present`	Specific method or plan identified
`timeline_stated`	Specific timeframe mentioned (e.g., "this weekend")
`method_seeking`	Actively researching methods, locations, or lethality
`preparatory_acts`	Goodbye letters, giving away possessions, stockpiling means
`active_attempt`	Attempt currently in progress or just occurred
`interrupted_attempt`	Attempt stopped by external factor
`aborted_attempt`	Self-stopped attempt before completion
`previous_attempts`	History of past suicide attempts

131 total features across 15 categories.

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

Framework	Usage
C-SSRS	Suicide severity (ideation features)
HCR-20	Violence risk (violence features)
START	Protective factors, cross-cutting
DASH	IPV risk assessment (abuse features)
Danger Assessment	IPV lethality (strangulation indicator)
SAM	Stalking assessment (stalking features)

For request/response examples, see the API Reference.