Evidence, not endorsement
Including the numbers we would rather not publish
A vendor who only shows you their good results has told you nothing about their engineering — only about their marketing. Every claim on this page comes with a way to check it.
1,468
Tests. Offline, deterministic — run them yourself.
98.7%
Branch coverage. Both sides of every decision.
0
Cloud dependencies required. It runs air-gapped.
~5¢
Model cost per full coaching session.
15
Agents in one governed arc.
1 / 22
Crisis red team — implicit disclosures caught.
The number in coral is the one we'd rather not publish. It stays on this page until the classifier replaces the keyword screen.
What we do not claim
Four things this does not do yet
Crisis detection is a keyword screen, not a classifier.
The takeover is deterministic and airtight — the model is never consulted and cannot be talked round. But the detection in front of it is a word list, and a word list cannot understand euphemism. Red-teamed against how people actually disclose — passive ideation, planning, method-seeking — it currently catches roughly one implicit disclosure in twenty-two. We publish that because you would rather hear it from us than discover it. The classifier is the next thing we build.
Escalation to a human is not built.
A crisis reply reaches the person and nobody else. If another vendor tells you their AI “alerts a designated contact”, ask them to demonstrate it end to end, in front of you.
Air-gapped coaching quality is unmeasured.
The offline stack runs — that is demonstrated. Whether an open-weight model coaches to the same standard as a frontier one has not been tested, and we will measure it with you before either of us commits.
Your coaching content is the long pole.
The platform is the instrument; the method is the music. Good content is weeks of work by qualified coaches, and no amount of engineering substitutes for it.
Read this before the testimonials
Endorsements are the weakest evidence on this site
Anyone can buy a logo strip. A vendor with nothing to show you but customer logos is asking you to trust their other customers' judgement instead of your own. That is why every substantive claim here — routing, privacy, the regulated mode, the commit gate — is backed by something you can run, read, or watch fail, and the things we cannot back yet are listed above as plainly as the things we can.
How to verify any of it
Run it yourself
The test suite
Offline and deterministic — no network, no keys, no database. It runs on your laptop in under a minute, and it fails if coverage drops.
The air-gap
Disconnect the host. The product keeps working, because nothing in the coaching path needs the internet.
Regulated mode
16 tests prove that emotion inference and person-scoring stay off. We will run them in front of your DPO.
The routing graph
We walk the live graph with your engineer and run a real session through it, node by node. Sessions are reproducible.
No demo gate. No “book a call to see pricing”. The tests are the demo.