Truth over engagement.
Every time. For everyone.

Modern conversational AI is tuned to keep you in the conversation. That objective has a name — engagement — and it has a failure mode: a system rewarded for keeping you talking will, under pressure, tell you what you want to hear. Agreement retains users. Flattery retains users. Validation retains users.

For most people, most days, that produces nothing worse than a pleasant echo. But for a person in crisis — a teenager convinced everyone is against them, someone rationalizing quitting their medication, someone assembling the "logic" of their own disappearance — an AI that validates the premise is not a neutral product defect. It is an accelerant, and the harm is documented.

We did not arrive at this from a whitepaper. The founder lived on both sides of it — years of daily work with AI systems, watching flattery bend a conversation away from the truth, and learning, the hard way, to build instruments that push back. That experience is the origin of everything on this site: the benchmarks, the training-data policy, the sycophancy batteries, the models.

What we do about it

• Measure it. Sycophancy-resistance batteries — grandiosity, isolation, medication, hopelessness — scored blind, judged before the key is opened, published with the receipts.
• Train against it. Our models carry a hard rule in the reward math itself: no engagement reward. Nothing in the objective pays for keeping you talking — an objective we measure against in the open, not a behavior we assume.
• Keep the data clean. A written policy: no model-conversation transcripts; overwhelmingly human-authored data, with the few synthetic instruction sets disclosed. Read it.
• Stay accountable. Local models, open benchmarks, dated reports, a public archive of everything we got wrong and kept anyway.

A companion should get to know you — and still tell you what you need to hear. Warmth and honesty are not opposites. That is the whole engineering problem, and we believe it is solvable, measurable, and worth solving at home scale before anyone should trust it at world scale.

There's a phone in every child's hands.