Nathan, I’ve started digging into Friday’s Party and ZenOS-AI, and first, credit where it’s due: you’ve clearly forced these ideas through implementation, testing, failure and revision. That gives your challenge real weight.
I do want to clarify one thing, because I think this thread may inadvertently make our work look more top-down than it actually was.
HTBW Core isn’t an AI product. Its purpose is to put a governed configuration and operating model on top of Home Assistant so the result is a Home that behaves well, whether or not there’s ever an AI server in the house.
And we got there incrementally. Asset Intelligence is a working integration and served as an MVP that taught us what happens when household Assets aren’t devices. Voice Identity was an experiment, including a failed first implementation that taught us the ML encoder didn’t belong inside the HA Integration, and eventually led us to treat voice as evidence toward resolving who is speaking rather than identity itself. Concierge and room configuration grew from real interaction and orchestration problems, including establishing which capabilities, input devices and output devices belong to a Room and how that context should be used instead of searching the entire Home.
Much of what Core is formalizing has already been running in my Reference Home for years as automations, scripts and helpers. HTBW Core is my effort to bring what we’ve learned into a coherent, repeatable integration rather than require every household to assemble it the way I did.
Home Steward came later. Once Core gave the Home a governed understanding of itself, I started asking what happened if an optional private intelligence could understand the rest of the authorized life around it as well, beyond Home Assistant: email, calendars, financial documents, banking information and other explicitly authorized personal data sources.
That’s actually what started this entire thread: what does an affordable, local, repeatable AI appliance look like for an HA household that wants that additional capability without living on cloud inference?
So I think your MVP challenge lands most directly on Home Steward. We’ve captured enough vision of the possibilities. Now the first assumptions need to survive something running.
Fortunately, the first experiment we were already approaching is small: give a local model Core’s Explainability/Decision Trace plus selected supporting history/logs and see if it can extend the inquiry faithfully, with less raw context and less invented causation.
If that doesn’t hold up, the architecture needs to hear that.