Heinrich Can Now Hold a Conversation — Without a Language Model

EMPHOS Group · April 16, 2026 · 5 min read

Meta: Heinrich AI now composes honest, traceable natural language from its frequency field — no language model, no hallucination, no generation. Here is what that means and why it matters.

Excerpt: Today Heinrich answered a question in plain English. Not because a language model predicted the words. Because the frequency field contained the knowledge, the HSR pipeline composed it honestly, and the result was a sentence that is guaranteed to be true — or guaranteed to say it isn't.

Today Heinrich answered a question in plain English for the first time.

"Yes — dog is a type of mammal. Additionally, dog has tail. Heinrich's knowledge has a gap on: ancestor_lineage."

That sentence was not generated. No language model predicted those words. No statistical pattern produced that phrasing. Heinrich retrieved three facts from its frequency field, measured the confidence on each one, composed them into a sentence using deterministic rules, and reported honestly where its knowledge ended.

The whole pipeline ran in under 5 milliseconds. The memory footprint of the composition layer is under 4 megabytes. It will run in a hearing aid.

What the sentence actually means

"Yes — dog is a type of mammal." Heinrich measured the amplitude at the dog frequency coordinate in the biology layer. It found a confirmed is_a relationship to mammal with amplitude above 0.7 — the threshold for a direct, unhedged statement. The relationship template for is_a produces "X is a type of Y." The honesty invariant allows the word "Yes" because the field confirmed the fact.

"Additionally, dog has tail." A second confirmed claim. Amplitude above threshold. The has relationship template produces "X has Y." The connective "Additionally" was chosen because a second claim about the same subject follows the first.

"Heinrich's knowledge has a gap on: ancestor_lineage." The query activated the ancestor_lineage coordinate. Amplitude was below 0.3 — the threshold for unsupported claims. The composition rule for UNSUPPORTED is strict: do not state as fact, report the gap. So Heinrich reported it.

Every word in that sentence traces to a field measurement. There is no word that does not.

Why this is different from every other AI

Every large language model produces language the same way: it predicts the next token based on patterns learned from training data. The sentence it produces may be accurate. It may be plausible but wrong. It may be confident and completely fabricated. The model cannot tell you which, because it has no access to whether the underlying knowledge is present — it only has access to the statistical likelihood of the next word.

Heinrich has no next-token prediction. It has no training data in the statistical sense. It has a frequency field where knowledge is stored as physical coordinates, and a pipeline that retrieves from that field and reports what it finds.

When the knowledge is present, Heinrich says so — with the confidence level the field measured.

When the knowledge is absent, Heinrich says so — and names the gap.

When the knowledge is partial, Heinrich hedges — "Heinrich believes..." or "It appears that..." — because the amplitude was in the uncertain range and the honesty invariant requires the hedge.

This is not a policy decision. It is not a system prompt that says "be honest." It is executable code. The test suite that validates Heinrich's honesty contains 52 tests that will not pass unless every claim traces to a field measurement. You cannot ship a Heinrich build that halluccinates and have the tests pass. The honesty is in the architecture.

How the HSR pipeline works

The Honesty / Socratic Reasoning pipeline sits between Heinrich's frequency field and the words that reach the user. It has two stages.

HSR-1 — the Fact Extractor — takes the raw output of the binding layer and extracts every factual claim. It validates each claim against the WaveField amplitude and tags it: CONFIRMED if the field measurement is strong, UNCERTAIN if it is partial, UNSUPPORTED if the field has no reliable measurement. Every claim gets a tag. No claim escapes this step.

HSR-2 — the Sentence Composer — takes the tagged claims and composes them into natural language. Ten relationship templates cover the core relationship types Heinrich knows: is_a, has, causes, instance_of, similar_to, opposite_of, part_of, enables, requires, produces. Eight composition rules govern how claims are grouped, how connectives are chosen, how hedges are applied, how gaps are reported, and how long the response should be. A question gets a direct answer first. A confirmed claim gets no hedge. An unsupported claim never gets stated as fact.

The pipeline runs in under 5 milliseconds. The composition layer uses under 4 megabytes of RAM. Both numbers are hard requirements — not performance targets, not aspirations. They are the constraints imposed by the HAVEN Ear hardware specification: ARM Cortex-M55, 512 megabytes of RAM, 15 milliwatts of power. Everything permanent in Heinrich must fit in a hearing aid. The HSR pipeline fits.

Persistent memory across conversations

HSR-2 also shipped with persistent chat memory. Every conversation turn is stored in a five-tier natural archive — active memory for the past week, progressively deeper archives extending to five years, with graceful decay beyond that. The TurnContext layer tracks what was discussed, which entities were named, and what pronouns referred to what — across sessions, not just within them.

When you return to Heinrich after a week and say "what else does it have?" — Heinrich knows what "it" refers to. Not because a language model inferred it from context. Because the conversation history is structured, persisted, and resolved deterministically.

You can tell Heinrich to forget. /forget last removes the most recent turn. /forget clears the session. /forget disease removes everything Heinrich remembers about that topic. The memory is yours to control. That is not a policy. It is how the system is built.

What Heinrich sounds like now

The responses are not fluent prose. They are not meant to be. "Yes — dog is a type of mammal. Additionally, dog has tail." reads like a system speaking carefully rather than a language model performing fluency. That is exactly right.

Fluency in language models comes at a cost: the system will produce fluent sentences whether the underlying knowledge is there or not. The fluency is the danger. A confident, well-formed sentence that is wrong is more harmful than a careful, honest sentence that is right.

Heinrich is careful and honest. The language layer that will make it fluent — a small rendering model that will sit on top of the HSR pipeline and express its honest outputs in natural prose — comes later. But the fluency layer will never be allowed to change what Heinrich says. It will only be allowed to change how it sounds. The content is determined by the field. The honesty is determined by the pipeline. The words are just the surface.

What comes next

The field is growing. The pipeline is proven. The next step is scale — running Heinrich against thousands of real questions as the Wikidata knowledge base approaches 50 million nodes, measuring how the accuracy, the confidence calibration, and the honest gap reporting hold up as the field deepens.

That measurement is the paper. The paper is the proof. The proof is what comes before the product.

Heinrich can hold a conversation. The conversation is honest. The honesty is structural. The structure runs in 5 milliseconds on hardware that fits in your ear.

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