PRODUCT 1 · LIVE
The Gate.
Before your model answers, know whether the answer is even in the context. One call returns a calibrated probability — so "don't answer" becomes a decision you can set, not a guess.
LIVE GATE
answer absentanswer present
The problem
Confidently wrong is the failure mode.
RAG answers even when the context doesn't contain the answer — fluently, and wrong. Watch it happen, then watch the gate stop it.
WITHOUT A GATE
“What's the early-closure penalty after year five?”
📄 Retrieved: clauses covering years one to three only.
⚠ fabricated — no clause supports this
WITH THE GATE
“What's the early-closure penalty after year five?”
📄 Retrieved: clauses covering years one to three only.
p(answerable) 0.12
Held — the answer isn't in the retrieved clauses. Routed to a specialist rather than guessing.
In the real world
What the gate catches.
Banking
Loan servicing assistant
“What penalty applies if I close after year five?”
⊘Not in the retrieved clauses — held, routed to a specialist.
Support
KB-backed deflection bot
“Why was I double-charged in March?”
⊘Account-specific — the KB can't answer. Escalated, not guessed.
Compliance
Policy copilot
“What's our data-retention window for KYC?”
✓Stated in the SOP — answered, with the citation attached.
Runs in your infrastructure
Nothing leaves your network.
The gate self-hosts. Your documents, your customers' questions, and the scores never touch our servers — same calibrated behavior as the hosted tier, deployed entirely inside your perimeter. Built for teams where data residency isn't optional.
The numbers
Measured, not asserted.
Benchmarks against LLM-judge baselines, calibration behavior, and self-hosted results — in the technical report.
Read the Gate technical report →