CalibratedAgentsBanking
Loan servicing assistant“What penalty applies if I close after year five?”
HELDNot in the retrieved clauses — held, routed to a specialist.
Gate · before your AI answers
Don't let it answer what it can't know.
Before your model writes a word, the gate checks whether the answer is even in what retrieval found. One call returns a calibrated probability — so “don't answer” becomes a decision you set, not a guess.
Worked example
Before generation
“What's my current interest rate?”
§4.2 Prepayment after 12 EMIs§4.3 Foreclosure charge 4%
0.07chance these clauses answer the question. Below your line — held, not guessed.
Try it: give the gate a question and some context.
Pick a question, drag in the clauses your retriever might find, and press Check. Leave the right clause out and watch the gate hold.
1Pick a question
2The loan agreement — drag clauses across
3What retrieval found · up to 3
P(answerable)0.00
Worked example, simulated in your browser with pre-scored clauses from a fictional loan agreement. The real gate scores any question against any context — talk to us to try it on yours.
What the gate catches.
Support
KB-backed deflection bot“Why was I double-charged in March?”
HELDAccount-specific — the KB can't answer. Escalated, not guessed.
Compliance
Policy copilot“What's our data-retention window for KYC?”
ANSWEREDStated in the SOP — answered, with the citation attached.
Self-hosted
Nothing leaves your network.
The gate self-hosts. Your documents, your customers' questions and the scores never touch our servers — the same calibrated behaviour as the hosted tier, inside your perimeter.
Talk to us about self-hosting →YOUR PERIMETER
Your appThe gateYour model
CalibratedAgents servers✕ NOTHING RECEIVED
Measured, not asserted.
Benchmarks against LLM-judge baselines, calibration behaviour and self-hosted results — in the technical report.
PDF · Gate technical report
How the gate is calibrated, what it was tested against, and how it runs self-hosted.Open the report →Questions, answered.
What does Gate do?
Before your model answers, Gate returns a calibrated probability that the retrieved context can answer the question. If it can't, the AI doesn't guess — a person answers instead.
Can I self-host Gate?
Yes. The gate self-hosts, so your documents, your customers' questions and the scores never touch our servers — with the same calibrated behaviour as the hosted tier.
How is Gate different from Ground?
Gate runs before generation and asks whether the answer is in your documents at all. Ground runs after and checks every claim in the answer against its source.
How was Gate tested?
See the Gate technical report and the Research page: trained on our own data factory, tested on public benchmarks and held-out documents, reported precision-first.
Give us a week of your logs. We'll show you what your AI got wrong.