CalibratedAgentsWrong value01
You can prepay after 6 EMIs 12 EMIs.
A number, date or amount differs.See an example →Ground · after your AI answers
Every claim, checked against the page.
Ground splits your AI's answer into claims, checks each one on its own evidence, and hands back a score, what's wrong, and the exact passage that proves it — before anything is sent.
What comes back · for a wrong claim
POST /v1/ground/basic
You can prepay after 6 EMIs 12 EMIs.
What's wrong
Contradicts the sourceWhy
The value changed: 6, not 12Where
Loan agreement §4.2RECEIPT · LOAN AGREEMENT §4.2Prepayment is allowed after 12 EMIs, subject to a foreclosure charge before year three.
Live today
values · scope · polarity · refusals · contradictionsPro · in development
recomputing numbers · following reasoning
values · scope · polarity · refusals · contradictionsPro · in development
recomputing numbers · following reasoning
Try it: give Ground the context, then check.
Pick an assistant, drag in the source passages, and press Check. Every claim is checked on its own — leave a passage out and the claim that needs it comes back not located.
QuestionAnswer
1Pick an assistant
2The documents — drag passages across
3Context Ground checks against
Ground's check
Worked example, simulated in your browser with pre-scored claims from fictional documents. For real scores on your own answers, use the playground in the console.
Not just “wrong” — marked up like an editor would.
Nine kinds of mistake · click one
Rolling out in the API
Wrong scope02
You can check in 30 kg on a domestic international flight.
True, but for another plan, person or place.See an example →Contradicts the source03
Room rent has no cap is capped at 1% a day.
Says the opposite of the document.See an example →Not in the source04
Annual health check-ups are free. no passage
Nowhere in the documents.See an example →Missing condition05
Part-payments are free‸ up to 25%, once a year.
Drops an “only if” or “up to”.See an example →Changed strength06
The insurer will may waive the co-payment.
“May” becomes “will”; a cap becomes a fact.See an example →Incomplete07
Covered: room rent and ICU‸ + ambulance, day care.
Lists only part of what the source lists.See an example →Sources disagree08
The wait is 36 months. §3.1 says 36 · the 2024 rider says 24
The documents conflict — not the AI's error.See an example →Needs reasoning09
Your EMI works out to Rs 8,560. PRO
A calculation or comparison — checked in Pro.See an example →The dial
You choose the false-alarm rate. Not us.
A judge gives you yes or no. Ground gives a score, so you pick the line: strict for a compliance queue, looser for an internal sweep — and you know what each setting costs.
0
of 6 wrong claims caught0
false alarms for your reviewers—
of flags are real
of 6 wrong claims caught0
false alarms for your reviewers—
of flags are real
20 ILLUSTRATIVE CLAIMS · WRONG ONES MARKED ✕ · OUTLINED = FLAGGED AT THIS LINE
One call. Receipts included.
Send the question, the context and your model's answer. If any step can't run, you get a 503 — never a silent pass.
curl https://api.calibratedagents.com/v1/ground/basic \
-H "Authorization: Bearer $CA_API_KEY" \
-d '{ "question": "What does my policy cover?",
"context": "…waiting period of 36 months…",
"answer": "Pre-existing diseases are covered after 24 months.",
"metadata": { "environment": "production" } }'
→ { "score": { "p_hallucinated": 0.97 },
"sentences": [{ "claims": [{ "text": "…covered after 24 months.", "status": "contradicted",
"evidence": [{ "text": "…waiting period of 36 months…" }],
"fix": "Correct it to what the source says: …36 months…" }] }] }
Questions, answered.
What does Ground check?
Each claim in an AI answer, against the context your system retrieved. Lookup claims — values, scope, polarity, refusals and contradictions — are checked today.
What does Ground return?
One calibrated score for the whole answer, and for every claim a status — contradicted, not in the source, depends on an error, or a reasoning step (Pro) — with the reasons, the source passage and a fix.
Does Ground check calculations and reasoning?
Not yet. Calculations, comparisons and conclusions come back marked as Pro reasoning steps; checking them is the Pro tier, in development.
How do I call Ground?
Send the question, the retrieved context and your model's answer to POST https://api.calibratedagents.com/v1/ground/basic, with an API key from the console.
Give us a week of your logs. We'll show you what your AI got wrong.