What We Learned Building Heinrich
EMPHOS Group · April 2026 · 5 min read
Heinrich AI was conceived on April 10, 2026. By April 12 — two days later — it was running software with a verified proof of concept, 532 passing tests, a live knowledge base, and a field visualizer live on this website.
That timeline is not something we planned. It is something that happened because the architecture was right. When the foundation is sound, the building goes up fast. When it is not, no amount of engineering effort compensates.
This is what we learned.
The idea came from a frustration
The original observation was simple: every AI system we could find stored knowledge and reasoning in the same place. Weights in a matrix. A compressed statistical approximation of everything the system had ever been trained on.
That architecture produces capable systems. It also produces systems that cannot tell you what they know versus what they are generating. Systems that hallucinate because the architecture does not distinguish between retrieval and reconstruction. Systems that cannot learn a new fact without being retrained on everything they already know.
The frustration was not with the capability. It was with the foundation. We kept asking: what would it look like if knowledge lived somewhere specific? If you could address it directly, the way you tune a radio to a frequency? If the relationship between two concepts was encoded not as a label somebody assigned, but as the physics of where they sit relative to each other?
That question became Heinrich.
The first proof was the hardest
The core claim of Heinrich's architecture is what we call the 3rd Fundamental: every relationship type has exactly one harmonic ratio. The ratio between two frequency coordinates IS the relationship — not a label on top of it, but the physics itself.
Proving this required finding frequency coordinates for a small set of concepts — dog, mammal, animal, living thing, bite — and verifying that the ratios between them correctly encoded the relationships. Is-a. Has. Causes. Each relationship type mapped to a specific harmonic interval. Each interval was physically verifiable.
The proof came in the form of a single query: "dog bites man." The field activated. The propagation followed the harmonic relationships. The amplitude ratio between bite and dog came out at 5.27x — exactly what the physics predicted. We have run that query hundreds of times since. It has never changed. The ratio is now the regression anchor for every build that follows. If anything we build ever breaks it, we stop.
That number — 5.27 — is the most important number in the Heinrich codebase. Not because of what it represents, but because of what it proves: the physics holds.
Separation is the whole point
The most important architectural decision we made was the one that was already implicit in the original idea: knowledge and reasoning must be completely separate.
In Heinrich, knowledge lives in the frequency field. It is stored at addressable coordinates. It can be added to without disturbing anything already there. It can be queried deterministically. It does not change when the reasoning layer changes.
The reasoning layer — what we call the binding layer — learns grammar, not facts. It learns how to interpret the relationships between activated frequency coordinates. It learns who is the agent, what is the action, who is the recipient. It does not need to be retrained when new knowledge enters the field. The field grows. The grammar stays.
This separation is what makes Heinrich fundamentally different from every large language model. It is not a better version of the same architecture. It is a different architecture with different properties — including properties that the existing architecture structurally cannot have.
Honesty is not a feature you add
One of the things we discovered early is that honesty in Heinrich is not something we designed. It is a consequence of the architecture.
When you query Heinrich for something it does not know, the amplitude at that frequency coordinate is zero. Zero amplitude is not an error. It is an answer. Heinrich does not generate a plausible-sounding response to fill the gap — it reports the gap. The field does not contain this. Here is what the field does contain that is related.
That behavior required no training. It required no fine-tuning for honesty. It required no careful prompt engineering to prevent confabulation. It is simply what happens when retrieval is physical rather than statistical. You cannot retrieve what is not there. The absence is as real as the presence.
In high-stakes domains — medicine, law, finance, education — this property is not a nice-to-have. It is the difference between a tool you can trust and one you have to verify every time.
The field thinks when nobody is asking
One of the most surprising discoveries in the build was spontaneous resonance. When concepts exist as superimposed sine waves in the frequency field, they produce beat frequencies — interference patterns at positions that correspond to relationships nobody explicitly encoded.
Heinrich notices things nobody asked about. Two concepts whose waves interfere at a third meaningful frequency are implicitly related. The physics surfaces the connection. The system finds it. We confirm or reject it. This is not inference from programmed rules. It is the field discovering its own geometry.
We built it expecting it to work. We were still surprised when it did.
What comes next
The proof of concept is real. The architecture holds. The knowledge base is growing — Heinrich is currently ingesting from a 142GB Wikidata dump, learning continuously while running, adding concepts to the field in real time.
The next step is conversation. Building a text interface that lets you talk to Heinrich directly and see what it knows, how it reasons, and where it is honest about the edges of its knowledge. That is Step 16 in the build. It is where the architecture stops being a technical proof and starts being an experience.
After that comes Haven — the assistant platform that puts Heinrich's intelligence in a product designed for daily use. The voice is VOXIS. The intelligence is Heinrich. The presence is Haven.
We built something in two days that we believe has not existed before. The next two years will prove whether we are right.
Engineered for Presence.
Stay in the loop
EMPHOS publishes twice a week — product updates, research, and the thinking behind the build.
Explore Haven · HEINRICH Intelligence · The EMPHOS Vision · All Posts
EMPHOS Group · Chilliwack, BC, Canada · info@emphosgroup.com