We Solved the Device Shipping Problem for AI. Here Is How.
EMPHOS Group · April 21, 2026 · 6 min read
Here is the shipping problem for any AI device: the intelligence layer — the knowledge field the device uses to think — is large. For HAVEN Ear, Heinrich AI's knowledge subfield for 500,000 curated concepts is approximately 88 megabytes as a working field array. That is manageable today. But as the subfield grows richer — more languages, more medical knowledge, more personal context accumulated overnight — it grows toward the storage limits of a device running at 15 milliwatts.
There is a more elegant question hiding inside that constraint: does the device actually need to store the full field at all?
The field can be regenerated from a compact manifest using only the harmonic grammar as a decompression key. Same input, same field, every time — deterministically, on any hardware, forever.
That insight became the Katz Protocol. This is the story of how we proved it.
Why the field is not the same as the knowledge
Heinrich AI stores knowledge as wave physics. Every concept has a unique frequency coordinate. Relationships between concepts are encoded as ratios within the harmonic grammar. The working field is a superposition of all these waves — efficient for inference, queryable in microseconds, but unable to reconstruct individual concepts from the waveform alone. The field is a summary. The manifest is the source of truth.
This creates a clean architectural question. If the harmonic grammar is a fixed constant — and it is, it never changes — then any device that knows the manifest and knows the grammar can reconstruct the full field deterministically. The field never needs to be stored. The manifest is the checkpoint. The field is the runtime.
The hypothesis was clean. But a hypothesis is not a proof.
The experiment
The Katz Protocol is a standalone research experiment with one purpose: determine whether lossless field reconstruction from a compact manifest is real, at what scale, and at what compression ratio. The experiment either confirms or falsifies the insight. Both outcomes are useful.
The pass condition was unambiguous:
max(abs(direct_field - reconstructed_field)) == 0.0
Not approximately zero. Not within a tolerance threshold. Exactly zero. The reconstruction is either lossless or it is not.
The experiment ran in five phases: write concepts directly to a field and store the exact array as ground truth; encode each concept as a minimal manifest record; reconstruct the field from the manifest using the harmonic grammar; compare; report every number honestly regardless of the result.
We ran it on four synthetic test sets — 100, 1,000, 10,000, and 100,000 concepts — then on real data.
The synthetic results
All four synthetic test sets passed. Lossless reconstruction confirmed from 100 concepts to 100,000. The harmonic grammar works as a decompression key at every scale tested.
The compression picture was more interesting than expected. The manifest grows linearly with concept count. The working field is a fixed size for a given frequency space, regardless of concept count. This creates a crossover point at approximately 58,774 concepts — below which the manifest is the compact form, above which the field is smaller. But the field cannot recover individual concepts. The manifest is always the lossless record. They serve different purposes, and both matter.
The real test — ConceptNet at one million concepts
Synthetic data proves a mechanism. Real data proves an architecture.
ConceptNet is Heinrich AI's first ingested knowledge source — 1,002,888 real English concepts and over 1.3 million real relationships, fully processed and mapped. Dog, bite, human, protein, gene, street, film — all real, all addressed, all carrying amplitude values shaped by the genuine structure of human knowledge.
We ran the full Katz experiment on all of them.
It passed. Lossless. The entropy impact factor of 1.07 is the number that matters most — real ConceptNet knowledge, with all its amplitude variation and real-world messiness, compresses through the harmonic grammar only 7% worse than clean synthetic data. The real world is remarkably well-structured when you look at it through the right lens.
What this means for HAVEN devices
HAVEN Ear's intelligence layer — 500,000 curated knowledge nodes — ships as a 4.3MB binary file. The device boots, loads the manifest, reconstructs the field in RAM. The entire operation takes seconds. Heinrich is awake. Query latency: two to twenty-one milliseconds. Server required: none.
Overnight, while HAVEN Ear charges in its bedside dock, the Socratic Engine identifies gaps encountered during the day and fills them from the full Heinrich field on the Home Hub. A delta manifest — only the new or updated concepts — is pushed to the device. New knowledge, silently, every morning. No reflashing. No firmware update. No subscription. No cloud.
The binary format that makes this possible stores exactly what is needed to reconstruct the field: Hz coordinate at 4 bytes, amplitude at 2 bytes, layer ID at 1 byte, phase at 2 bytes. Nine bytes per concept. At that density, HAVEN Ear's full 500,000-node subfield fits in approximately 4.3 megabytes. A full million-concept field fits in under 9 megabytes. Every HAVEN device carries its complete intelligence layer in a file smaller than a handful of photos.
The shipping problem is solved.
The deeper implication
Katz was scoped as a compression experiment. What it proved is something more fundamental about what Heinrich AI is.
The field is ephemeral. The manifest is permanent. The harmonic grammar is the key that turns one into the other.
This means the knowledge base is not a large binary that must be managed, versioned, and shipped with every device update. It is a compact structured record that can be reconstructed on demand, on any hardware that can run the reconstruction. That is not a storage optimisation. It is a statement about what kind of thing Heinrich fundamentally is — and why it can live in a hearing aid, a pair of glasses, a watch, or a handheld clinical device, all running the same intelligence from the same source of truth, updated silently every night.
What comes next
Three questions remain on the Katz research agenda. Does standard compression applied to the binary manifest yield another meaningful size reduction — the manifest has predictable structure, and Hz values cluster around harmonic positions. Does lossless reconstruction hold at Wikidata scale — ConceptNet at one million passed; the full Heinrich staging store now holds 68 million concepts and is growing. What is the RAM requirement for the reconstructed field on HAVEN Ear's target hardware.
These have answers. We will measure them.
The Katz Protocol is now part of the HAVEN device shipping specification. Every HAVEN device will boot from a binary manifest. The field will never be stored as a static file. Heinrich travels light.
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