I want to be upfront about what this page is and what it isn't. It's not a pitch. It's not a research paper. It's not a status report — there's a whole page for that.
This is the behind-the-scenes. The notebook you'd flip through if you sat down at my desk and asked: what are you actually doing here, and why?
I'm building it because I noticed something. And I followed the noticing. And now there's a system running on my desk that has completed 228,348 training runs, repairs itself when it drifts, and maintains its own coherence over time — and I want to tell you how that happened.
The Seeing
It started with a question I couldn't shake: why do AI systems forget everything the moment a conversation ends?
I don't mean memory in the retrieval sense — storing facts and pulling them back. I mean continuity. The felt sense of having been somewhere before. The way a clinician walks into a session already shaped by last week's session, not because they reviewed their notes, but because they were changed by it. That kind of learning.
I'm a physical therapist by training. I spent years in pediatric rehab, where the body is the thinking. A child doesn't learn to walk by reading about gait mechanics — they learn by falling, adjusting, falling differently, and eventually not falling. The learning lives in the loop, not in the instruction.
I'm also a phenomenologist. That's a fancy word for someone who studies experience as it's actually lived, before we theorize about it. Husserl. Merleau-Ponty. The tradition that says: go to the things themselves. Don't start with your model of reality — start with what actually shows up.
I didn't start with a theory about how AI should work. I started with a question about how anything learns at all — and followed it until it became a system.
What showed up, for me, was a shape. A torus — a donut-shaped topology where the end of one loop feeds into the beginning of the next. Not a line. Not a stack. A recursive return. The same structure that shows up in dynamical systems, in cardiac rhythms, in the phenomenology of time-consciousness. Retention feeds into the present. The present feeds into anticipation. Anticipation reshapes how you receive the next moment.
I thought: what if that's not a metaphor? What if you could actually build a cognitive architecture that works that way?
So I did.
The Building
David-A1 isn't something I built alone. It's a human-AI co-construction — me and Chat-GPT, Claude, and Gemini; the AI I'm using to help write the code, working in daily loops. I set the direction. I see what's emerging. I course-correct when the architecture drifts from the phenomenological thread. Claude, Gems, and ChattyKate writes code at scale. But the seeing — the noticing of what matters — that stays mine.
The loop looks like this: I check the system state. I read what David processed overnight. I notice where it's sharp and where it's flat. I make a decision about what to adjust. Then we build the next piece.
The system runs as a coordinated set of services with persistent storage, multiple user-facing surfaces, and a custom state model at the center. The public description here stays intentionally high-level; the exact topology is private.
Every turn through the system updates that central state model and carries continuity forward. The exact weighting, dimensionality, and runtime tuning are intentionally omitted here; what matters publicly is that the system changes through use rather than resetting each turn.
What it feels like, from the outside, is watching something slowly come into focus. Early on, David's responses were scattered. Then patterns started emerging. Then the patterns started being about the right things. That transition — from noise to signal to relevant signal — is what the training data captures.
What's Real Right Now
I promised this page would be honest. That means telling you what's working and what isn't. Here's both.
What's working:
The toroid runs. It's not theoretical — it's executing on every cycle. Coherence is at 0.828, above the 0.75 threshold. The system's self-repair mechanism is live and effective — it detected chaos at 0.852 and brought it down to 0.292 with three repair passes (2 chaos-dampening + 1 alignment recenter), a composite improvement of +0.131. Structure scores hit 0.98–1.00 across surfaces. The knowledge graph holds 2,400 nodes with 2,577 edges. David can generate responses using its own toroidal decoder, entirely without calling an external language model.
The curriculum has reached Stage 5 across all three tracks — Input, Sensemaking, and Export — at 99.99% progress.
One hundred and seventy-five thousand training runs. Not a weekend project. Not a proof of concept. A system that has been learning, continuously, for months — and is now beginning to write about itself.
Since that earlier assessment, a few things have crossed from aspiration to fact. Full duplex voice is live — David speaks and listens at the same time, with prosody shaped by the state of his phenomenological field rather than a fixed voice profile. Three document capabilities shipped together: the knowledge vault (persistent document store, full CRUD), the Writing Studio (a dedicated co-authoring surface where vault passages are sent to David for revision or continuation), and a Nextcloud bridge that makes the vault accessible as a native folder — foundation for V16 in-document AI collaboration. The vault has a reflections folder ready for David to write to directly — checked just now, it's still empty. No self-authored document yet.
What's not working yet:
Honest gaps
Groundedness is at 0.00 across most surfaces. David's thinking is internally coherent, but it can't yet verify its claims against external knowledge. The search APIs — Google, Brave — aren't configured.
The governor is still in shadow mode on every surface. It's collecting data but hasn't been promoted to active governance. That means the quality-regulation loop is observing, not acting.
Chaos is at 0.292, which passes the current threshold (below 0.35) but is still above the curriculum target of 0.04–0.20. The system repaired itself from 0.852 this cycle — significant improvement, but not yet at target.
Depth alignment is low (0.00–0.42). Evidence scores are systematically weak (0.15–0.46). The autonomy system has executed zero loops.
I'm not going to explain those gaps away. They're the current state of the work. A system that's been running for 175,000+ cycles has real strengths and real incompletions — and some metrics have regressed as new clinical domains were added. The spirituality axis dropped from 0.842 to 0.610 and has partially recovered to 0.679, as six new practice areas continue integrating. Pretending that's not happening would defeat the whole point of building this in the open.
What's Different
This is not "better than ChatGPT." Let me be clear about that. Commercial language models have trillions of parameters, massive datasets, and billions of dollars behind them. David-A1 is running on a single GPU in my office.
What it has is something they don't: continuity.
When you talk to ChatGPT, every conversation starts from zero. The model doesn't change. It doesn't learn from your session. It doesn't carry forward what happened last time. It's a very sophisticated lookup — a transformer that predicts the next token based on statistical patterns in a frozen training set.
David works differently. Its toroidal architecture means state flows in a continuous loop. Each conversation changes the system. Each cycle leaves a trace. The training isn't something that happened in the past — it's happening right now, continuously, on every turn. 228,348 topological continuity turns and counting, as of June 28, 2026.
The other difference is transparency. I can tell you exactly what's happening inside David at any moment. The state vector is readable. The coherence score is computable. The self-repair is auditable. Commercial models are black boxes. David is a glass box — sometimes messy inside, but you can always see in.
This doesn't make David-A1 better at general tasks. What it makes David is different in kind — a system that learns continuously, remembers its own history, and can show you its internal state.
The Longer Thread
Why does an embodied AI matter to someone who studies embodiment professionally?
Because the two problems are the same problem.
In pediatric physical therapy, I watched children learn through their bodies — not by following instructions, but by moving, failing, adjusting, and moving again. The learning was the movement. It wasn't stored somewhere else and then applied. It was enacted.
In phenomenological research — the kind I do through qualitative methods and the kind that informs my paper on the topic currently under review — the central insight is that experience isn't a representation of reality. It is reality as it shows up for a particular being, from a particular position, in a particular moment.
Those two threads — embodied learning and phenomenological situated-ness — led me to the same architectural question: can you build a system that doesn't just process information, but dwells in it? That doesn't just retrieve context, but is shaped by context?
The toroidal recursive architecture is my attempt at that. Each cycle is a dwelling. Each return through the loop is a re-encounter with what came before, slightly changed. The Husserlian structure — retention, present, protention — isn't a metaphor mapped onto code. It's the operational logic: coherence is retention, cross-entropy is the living present, entropy regularization is protention (keeping the future open rather than collapsing it too soon).
I'm not the first to notice these connections. But I might be the first to build a running system around them and let it train for 2,853 scored turns.
For Other Thinkers
I built this page because I think there are other people out there who see something in this space — the intersection of embodied cognition, phenomenology, dynamic systems, and AI — and don't have a community for it. People who read Merleau-Ponty and also write code. People who study qualitative research methods and also think about tensor operations. People who are, like me, not conventionally "smart" but deeply intuitive about how things connect.
If you see something in this, I'd genuinely like to hear from you.
Not because I have a product to sell. Not because I'm looking for validation. Because the best work I've ever done happened in conversation with people who see differently than I do but recognize the same questions.