Byproducts of curiosity that solve real problems.

You and your partner disagree about the thermostat. Your roommate has opinions about the lights. Your teenager has a different sleep schedule. In most smart home systems, someone wins and everyone else loses. MANAS treats this as what it actually is: a multi-agent decision conflict with uncertain preferences.

Under the hood, Parallax handles the formal argumentation. Personality profiles become arguments with different levels of certainty. Environmental context, like the weather outside, enters the reasoning as real-time data. The result is not a compromise. It is a formally justified resolution that accounts for how strongly each person actually feels.

The intelligence runs locally. Your household data stays in your house.

An honest note on sequencing. MANAS has pivoted shop first. The retail and branding surface comes first, and the reasoning engine activates behind it, rather than the other way round. That is a deliberate reversal of the original plan, and we would rather say so on this page than quietly ship a different product than the one we described.

BUDDY is a small, round, friendly robot designed for children in Montessori environments. It looks like something a child would want on their desk. The creature on its screen has a species, a personality, and a growth arc tied to the child's own learning journey.

The species is not random. It is determined by a method unique to each child, revealed through a workshop experience that builds anticipation over weeks. The behavioral assessment running underneath uses play-based games, not clinical tests. Children engage naturally. The data emerges from how they play, not what they report.

Where this actually stands: BUDDY is a concept in design, not a running pilot. Nothing is deployed in a classroom and no child has used one. The voice layer, when it activates, will pass every interaction through our governance pipeline, and three guardrails are fixed before a single unit ships: no biometric signals, nothing ever framed to a child as emotion, and no consequential decision about a child driven by this system. If you are a Montessori educator who wants to shape the pilot before it exists, that is exactly the conversation we want.

Not every project on this page is ours alone to describe. This one is a working engagement with a clinical partner, and the parts that belong to them stay theirs to announce. What we can describe is our own side of it, and the honest reason to describe it at all is that the most useful thing that happened here was a rejection.

We proposed a personalization layer. The client declined it. That is the whole event, and it is not a disaster: a client who says no to one component and yes to the engagement is telling you something specific and true about what they actually need.

What we did with the no. We did not rename the declined component and put it back on the table under a friendlier description. Relabeling a rejected feature to slip it through is a failure mode we hold a standing rule against, because the new description is almost never a true account of what the thing does. Instead the rejection forced a reframe of the engine itself, and that reframe is the most valuable thing this lab got out of the first half of 2026. We would not have reached it on our own. The fuller account is in The Threshold, Issue 01.

If you run a clinic and want to know what this looks like in practice, ask us and we will tell you what we can.

CLAW exists because we needed to guarantee that AI interactions with children would be safe. Not marketing-safe. Formally verifiable safe. So we built a multi-stage pipeline where sensitive information gets caught, policy constraints get enforced, and argumentation-based conflict resolution handles the edge cases that rule-based systems miss.

The pipeline is a companion to Parallax, not a subsystem of it. It has its own architecture, its own scope, and its own repository, and it is the one piece of infrastructure we co-created with Saatvix, a cybersecurity company and a separate organization from this lab, which is why it is built to be inspected by people who did not write it. When BUDDY Bot's voice layer eventually speaks to a child, CLAW is what will stand behind what it says. The argumentation layer uses the same formal methods as Parallax, applied to the specific challenge of governing AI output in sensitive contexts.

It is open source. We think AI governance tooling should be inspectable by anyone who cares enough to look.