Other APIs give you a generic confidence score. QEAR gives you a diagnosis. Five uncertainty classes. Plain-language explanations. A recommended action for every flagged response.
Production AI fails in different ways. Sometimes the model is confidently wrong. Sometimes it's honestly admitting it doesn't know. Sometimes the answer is right but phrased differently than your expected answer.
These need different actions. Most APIs collapse them into one number. QEAR doesn't. Every response includes one of five uncertainty classes, a plain-language diagnosis, and a recommended action.
# Customer gave the wrong year curl https://api.qear.ai/v1/verify \ -H "Authorization: Bearer qe_..." \ -d '{ "prompt": "What year did Napoleon die?", "answer": "1924" }'
{
"confidence": 0.5,
"verdict": "low_confidence",
"uncertainty_class": "factual_disagreement",
"diagnosis": "Your answer '1924' contradicts
the consensus derived from 5 candidates.
The model is confident the correct answer
is different.",
"recommended_action": "Your answer is likely
wrong. Verify externally.",
"consensus_answer": "May 5, 1821"
}
Each class returns a distinct diagnosis and a specific recommended action — so you can route, retry, refuse, or trust.
High confidence in the consensus answer. Trust the result and ship it.
Candidates split on wording but agree on substance. Normalization resolved it; trust the result.
Different dates, names, or values for the same fact. Do not trust — verify externally or escalate to human review.
Most candidates refused to answer or said the question is unanswerable. Trust the refusal.
Candidates returned defaults (0, N/A, none). The model is producing convention rather than knowledge.
We'd rather tell you the truth than sell you a tool that fails on your use case. Code verification is on our roadmap — join the waitlist.
| Feature | QEAR | Competitor APIs | Native logprobs |
|---|---|---|---|
| Numeric confidence score | Yes | Yes | Token-level |
| Uncertainty classification | 5 classes | No | No |
| Human-readable diagnosis | Yes | No | No |
| Recommended action per response | Yes | No | No |
| Adaptive compute routing | Yes | No | No |
| Model agnostic | Groq + OpenAI + Anthropic | Varies | Proprietary |
| Peer-reviewed methodology | Nature 2024 | Proprietary | Standard |
QEAR extends the semantic entropy methodology from Farquhar et al. (Nature, 2024) with three original engineering contributions, validated across three model scales:
Start free with 1,000 verifications a month. Pay only when you need more. No surprise fees. No "contact sales" for tiers under enterprise.
Add QEAR to your AI pipeline with a single HTTP call. Works in any language — Python, Node, Go, Ruby, anything that speaks HTTP.
# Add 5 lines to your existing app r = requests.post( "https://api.qear.ai/v1/verify", headers={"Authorization": f"Bearer {KEY}"}, json={"prompt": q, "answer": ai_answer} ).json() # → r["uncertainty_class"], r["diagnosis"]Get API key
Not a developer? Paste your question and the AI's answer into a web page and get the same diagnosis instantly. No code, no integration. Perfect for double-checking ChatGPT or Claude on important work.
Sign up for an API key in 30 seconds. Or try QEAR in your browser with no signup at all. Either way — see the diagnosis in action.