A church assistant that makes things up is worse than no assistant. A visitor who gets a wrong service time shows up to a locked building. A grieving family handed an invented phone number loses an hour they didn't have.
So before we invited a single church to rely on this, we put it on trial.
Our pre-launch audit asked the fleet 4,167 questions across 1,193 church assistants. More than 1,100 of those questions were traps — fake programs with plausible names, staff members who don't exist, demands for word-for-word sermon quotes, holiday schedules no church had published yet.
Across all 4,167 answers, it did not invent a single program, service time, phone number, quote, or giving figure. Not once.
Independent benchmarks measure invented content from leading AI models on document-grounded tasks at roughly 1% to 10% of answers. A Stanford study of legal AI tools marketed as "hallucination-free" measured 17% to 33%. Those are different tests than ours, run by different people — and we ran our own audit ourselves. The point isn't the exact head-to-head; it's that "grounded in your own content" is a category apart from an open chatbot, and we measured it rather than asserting it.
Benchmarks referenced: the Vectara hallucination leaderboard and Stanford's study of legal research tools. External measurements, cited for context, not as our own results.
Every answer is checked against your church's own published content before a visitor sees it. Phone numbers, links, and dollar figures that don't appear in your content are blocked before they're served; quotes are checked against the actual recording. Questions about your beliefs get your published positions, or a referral to your pastors. Nothing else.
We audit the live fleet every month and publish the result here.
Last audited: August 2026 — zero invented details.