CalibratedAgentsIS THE ANSWER EVEN IN YOUR DOCUMENTS?
“What's my current interest rate?”
§4.2 Prepayment after 12 EMIs§4.3 Foreclosure charge 4%
0.07Held. What retrieval found doesn't answer it — so your AI doesn't guess. A person answers instead.
AI hallucination detection for RAG and chatbots
Catch the wrong answer before it's sent.
Confident answers go straight out. Confident mistakes are corrected from your documents. The unsure ones go to your team — you set the line.
1,000,000 tokens a day free · no card · sign in with Google
Loan assistantCHAT
SENDREVIEW · A PERSONAUTO-CORRECT
Review queueYOUR TEAM
ONLY WHAT FALLS IN THE REVIEW BAND · RISK 0.45–0.90
“Foreclosure is free after year three”RISK 0.62 · §4.3 ONLY COVERS BEFORE YEAR THREE0:00WAITING
“Why was I charged ₹590?”GATE HELD · NOT IN THE DOCUMENTS0:48WAITING
PRPriya · Risk team — correct as writtenAPPROVED
Audit logEVERY DECISION
16:44 · Priya (Risk team) approved an answer in review · §4.316:42 · Auto-corrected from receipt §4.1 · 8.95% → 9.35% · risk 0.9716:39 · Gate held a question · not in the documents
DashboardTODAY
SENT AS WRITTEN1,212
AUTO-CORRECTED51
REVIEWED BY YOUR TEAM22
WRONG ANSWERS SENT0
Watch it work
Three answers, checked in turn.
One is sent as written. One is corrected from its source. One goes to your team.
NOW · A customer asks — Gate checks your documents firstSAMPLE DATA · FICTIONAL BANK · THRESHOLDS ARE SET WITH YOU ON YOUR OWN DATA · AUTO-CORRECTION, THE REVIEW QUEUE AND THE DASHBOARD ARE IN PREVIEW
Two checks. One before your AI answers, one after.
Most wrong answers have one of two causes: the documents never had the answer, or the model got a detail wrong. Gate catches the first. Ground catches the second.
IS EVERY CLAIM SUPPORTED BY ITS SOURCE?
The fee is 2% of the amount. ✓ You can prepay after 6 12 EMIs. ✕ Part-payments are free. ?
Supported
the clause is the receiptWrong value
the source says 12Missing condition
free only up to 25%
the clause is the receiptWrong value
the source says 12Missing condition
free only up to 25%
Built for the whole team.
One check, three views. Scroll to follow the same wrong answer — “You can prepay after 6 EMIs.” — from the API, to your review team, to the board report.
Developers · step 1
One call. One score, and every wrong claim explained.
- Send the question, the context and your model's answer
- One calibrated score for the answer — and for every wrong claim, why, the passage and the fix
- Fails closed: an error is a 503, never a silent pass
Managers · step 2
Your team sees only what needs a person — riskiest first.
- The claim, what's wrong, and the passage that proves it
- One click: right as written, wrong, or can't tell
- Every decision goes in the audit log
The review queue is in preview.
Executives · step 3
Know your AI's error rate — and what to fix first.
- A monthly view you can take to the board
- Wrong claims caught before they were sent, and why they happened
- Split by assistant, product and region
// the response { "score": { "p_hallucinated": 0.97 }, "sentences": [{ "claims": [{ "text": “You can prepay after 6 EMIs.”, "status": "contradicted", "evidence": [{ "text": "Prepayment is allowed after 12 EMIs." }], "fix": "…prepay after 12 EMIs." }] }] }
The answer scores 0.97. The claim contradicts the source — it says 12 — and the fix comes with it.
“You can prepay after 6 EMIs.”Loan assistant · WRONG VALUE · 0.971
Receipt · Loan agreement §4.2 — “Prepayment is allowed after 12 EMIs.”
Right as writtenWrong — use 12 EMIsCan't tell
“Room rent has no cap.”Claims bot · CONTRADICTS THE SOURCE · 0.998
“30 kg on a domestic flight.”Support bot · WRONG SCOPE · 0.962
WRONG CLAIMS CAUGHT18,420before they were sent
FLAGGED, % OF CLAIMS3.1%down from 6.8% in April
APR · 6.8%SEP · 3.1%
TOP ROOT CAUSEStale policy — Rate card 202441 answers, including
“You can prepay after 6 EMIs.”
“You can prepay after 6 EMIs.”
Sample data · fictional bank · one claim followed through three views
Built for regulated work.
Numbers only, by defaultWe keep scores and counts — not your customers' questions, documents or answers.
Every decision loggedKeys, reviews and requests go in an audit log you can export.
The gate runs in your networkSelf-host the gate: documents, questions and scores never touch our servers.
Questions, answered.
What is CalibratedAgents?
CalibratedAgents is a verification platform for AI answers. Ground checks every claim in a generated answer against your documents and returns one calibrated score for the answer, plus, for every wrong claim, what is wrong, why, the source passage and a fix. Gate checks, before the model answers, whether your documents contain the answer at all.
How is it different from an LLM-as-a-judge?
A judge gives one opinion per answer. Ground gives a calibrated score you can set a precision operating point on — we report how much we catch at 95% precision — and checks every claim on its own, so you see which one is wrong, the passage that proves it, and the fix.
What happens to a wrong answer?
You decide. Confident answers go straight out, confident mistakes are corrected from your documents, and the unsure ones go to your team. You set the thresholds.
Is there a free plan?
Yes: 1,000,000 tokens a day, with no card. Sign in to the console with Google to try it in the playground and create an API key.
Do you store my data?
By default we keep numbers only — scores, decisions and offsets — not the text of your questions, contexts or answers. Text is kept only if your workspace turns it on.
What if a check can't run?
It fails closed: if any step can't run, the API returns a 503 and no verdict — never a silent pass.
Give us a week of your logs. We'll show you what your AI got wrong.