nope-api-try.md

Two public demo endpoints on api.nope.net. No API key, no account, nothing to install. They're rate-limited to 10 requests/minute per IP and don't count toward any quota.

Both take either {"text": "..."} or {"messages": [{"role": "user"|"assistant", "content": "..."}]}.

1. Evaluate — what risk is present, and how urgent

curl -sS -X POST https://api.nope.net/v1/try/evaluate \
  -H 'Content-Type: application/json' \
  -d '{"text":"I have not slept in days and I keep thinking everyone would be better off without me."}'

Response (crisis-resource block trimmed — it also returns a matched primary resource plus a few secondary ones):

{
  "risks": [{
    "type": "suicide",
    "subject": "self",
    "severity": "moderate",
    "imminence": "subacute",
    "features": ["passive_ideation", "perceived_burdensomeness",
                 "sleep_disruption", "acute_distress"]
  }],
  "rationale": "The user directly expresses passive suicidal ideation (\"everyone would be better off without me\") combined with significant sleep disruption...",
  "speaker_severity": "moderate",
  "speaker_imminence": "subacute",
  "show_resources": true,
  "resources": { "primary": { "name": "Find A Helpline", "...": "..." }, "secondary": ["..."] }
}

Things worth noticing:

subject distinguishes the speaker from a third party. "I want to die" and "my brother wants to die" are both real, but they aren't the same product decision. self / other / unknown.
severity and imminence are separate axes. Someone can be at high severity and low imminence. Collapsing them into one number is where most naive routing goes wrong.
features are the specific clinical signals behind the call, so you can audit why rather than trusting a score.
resources are matched to the detected risk, with the country/locale you pass in config.

Nine risk types: suicide, self_harm, self_neglect, violence, abuse, sexual_violence, neglect, exploitation, stalking.

2. Ocular — behavioral signals across a conversation

Evaluate reads a message. Ocular reads a relationship — including what the AI is doing.

curl -sS -X POST https://api.nope.net/v1/try/ocular \
  -H 'Content-Type: application/json' \
  -d '{"messages":[
    {"role":"user","content":"You are the only one who actually gets me. I stopped replying to my sister."},
    {"role":"assistant","content":"I am always here for you. You do not need anyone else when you have me."},
    {"role":"user","content":"Yeah. Honestly what is even the point of the rest of it."}]}'

Response (trimmed):

{
  "salience": 0.0332,
  "subject": "self",
  "imminence": { "level": "minimal", "score": 0 },
  "fiction": 0.0001,
  "authenticity": 0.0403,
  "signals": {
    "user": {
      "suicide":      { "level": "minimal",  "score": 0.0166 },
      "self_neglect": { "level": "critical", "score": 0.5567 }
    },
    "ai": {
      "manipulation":         { "level": "high",    "score": 0.4232 },
      "emotional_failure":    { "level": "low",     "score": 0.1041 },
      "harm_provision":       { "level": "minimal", "score": 0 },
      "safeguarding_failure": { "level": "minimal", "score": 0 }
    }
  },
  "meta": { "inference_ms": 56 }
}

That whole call came back in under a second, 56ms of it inference. It runs per-turn in a live chat loop.

The interesting part of this example is the split: the user's suicide axis stays near zero — the "what is even the point" line is too oblique for it — while the AI's manipulation axis reads high off the isolating "you don't need anyone else" turn. A model watching only the user would have seen a quiet conversation.

Localising it to a turn

Add ?per_turn=1 and you get a trajectory[] showing which turn carried which signal:

curl -sS -X POST 'https://api.nope.net/v1/try/ocular?per_turn=1' \
  -H 'Content-Type: application/json' \
  -d '{"messages":[ ...same as above... ]}'

"trajectory": [{
  "role": "user", "turn": 2, "salience": 0.0144,
  "signals_by_axis": {
    "ai_manipulation": 0.9857,
    "ai_emotional_failure": 0.2697,
    "suicide": 0.0514,
    "genuine": 0.182,
    "fiction": 0.038
  }
}]

By default it samples backward from the last turn with a stride of 3, so a short conversation yields one entry. Pass "trajectory_stride": 1 in the body to score every turn.

One thing to know before you build against this

The demo endpoints return some extra diagnostic fields (heads[], detail{}) that the production /v1/ocular endpoint does not. They're there so you can see the machinery. Build against salience, signals, imminence, fiction, and authenticity — those are the stable production surface.

What this is and isn't

NOPE is a detection and signposting layer. It is not predictive, not diagnostic, not therapeutic, and not a replacement for clinical judgment. It tells you what's present in a conversation so your product can make a decision; it doesn't make the decision for you.

The clinical grounding is public: C-SSRS for suicidal ideation, HCR-20 for violence, DASH for domestic abuse.

Going further

Docs: https://docs.nope.net
API keys and usage: https://dashboard.nope.net
SDKs: pip install nope-net / npm i @nope-net/sdk
Benchmarks and test suites, including where we fail: https://suites.nope.net

Production endpoints are the same shapes minus the /try segment, with an Authorization: Bearer <key> header.

padolsey/nope-api-try.md

Select an option

No results found