Heinrich and the Large Language Model. A Head to Head

EMPHOS Group, May 30, 2026, 7 min read

Heinrich and a large language model can both answer the question "what is a mammal?" Both return a coherent English sentence. To a person looking at the screen, the two outputs are interchangeable.

The physical work that produced them is not.

An LLM produced its answer by running a forward pass through several hundred billion learned weights, predicting the most probable continuation of your question token by token, then stopping when it had something that looked complete. Heinrich produced its answer by resolving the word "mammal" to a six dimensional address in a harmonic field, striking that address, reading the activation cloud that rang in response, and composing a sentence that traces every clause to a source claim.

The argument for Heinrich has never been that it is a better LLM. It is that it is not an LLM at all. The two systems share an output category and almost nothing else. This post is the head to head, line by line.

How knowledge is stored

An LLM stores knowledge as numerical weights. Hundreds of billions of floating point values arranged in attention layers and feed forward networks. There is no row anywhere in an LLM that says "a mammal is an animal." That fact is distributed across the weights, encoded as a statistical pattern absorbed during training. You cannot point at it. You cannot edit it. You can only retrain the model and hope the new training run preserves it.

Heinrich stores knowledge as positions in a harmonic field. The fact "a mammal is an animal" is the harmonic ratio between two specific frequency coordinates. An octave. 2:1. The canonical interval that means "is a" in Heinrich's geometry. The ratio carries the meaning. The physics of the field is the semantics. There is no separate label saying "this edge is an is a relationship." The interval between the two addresses already says it.

You can point at the row. You can edit the value. You can add a new fact without disturbing anything else in the field.

How a question becomes an answer

An LLM does not retrieve. It reconstructs. Every response is generated token by token from the statistical patterns the model absorbed during training. The model is not looking anything up. It is predicting what plausible continuation should come next. When the prediction lands on a correct fact, that is because the pattern for that fact was thoroughly present in the training data. When the prediction lands on a wrong fact that sounds correct, that is because the pattern looked similar enough. The model has no way to distinguish between the two cases.

Heinrich retrieves. A query resolves to a set of seed concepts. Those seeds are struck at their addresses. Energy propagates through the field along three pathways. Along the same harmonic line through canonical bias ratios. Across lines through composite signature coupling. Across layers through root frequency ratios. The activation cloud that results is the answer. Same query, same cloud, every time. Deterministic by construction.

Retrieval either reaches an address or it does not. There is no middle ground where the system invents something plausible to fill the gap.

How the system learns

An LLM cannot learn a new fact in production. The weights are frozen at training time. To add knowledge to the model, you have to do one of three things. Include the new fact in the next training run, which means months of work and millions of dollars. Fine tune on a small dataset, which destabilises every other capability the model has. Or graft on a retrieval augmented layer that fetches text from an external database and stuffs it into the prompt, which is a workaround, not a fix.

Heinrich learns by writing to an address. A new concept is placed at a coordinate. A new relationship is the ratio between two coordinates. The field grows. Everything already in it is undisturbed. There is no retraining cycle, no fine tuning, no fragile workaround. The system that answered "what is a mammal" at 9 a.m. answers it the same way at 9 p.m., except the cloud activated is richer, because more related concepts have entered the field in between.

Heinrich is learning right now, on a laptop, while it is being queried. The Wikidata ingest pipeline streams claims into the substrate continuously. The field grows. The mind keeps working.

How honest the system is about what it knows

An LLM cannot tell you what it knows. It can tell you what it produces. When it answers a question correctly, you do not know whether that is because the model genuinely absorbed the fact during training or because it generated a plausible sounding response that happens to be right. When it answers incorrectly, you have no way to tell from the answer itself. Hallucination and accuracy look identical from the outside. They have to be verified externally.

Heinrich's honesty is structural. If a query resolves to a concept the field contains, the cloud activates and a grounded answer comes out. If the query resolves to a concept that does not exist in the field, or if the propagation cannot find a supporting path, the system reports the gap. "The field knows about disease but found no strong connections for this query." Not "I think the answer might be." The field actually does not contain this. Zero amplitude at the coordinate is not an error. It is a measurement.

On the benchmark suite Heinrich runs against today, including simple facts, unknown entities, and ambiguous queries, the hallucination rate is 0 percent across every category. The system does not produce a claim it cannot trace. By rule. By architecture. By physics.

What it costs to run

A frontier LLM runs on warehouse scale infrastructure. Each inference step executes a forward pass through hundreds of billions of parameters streamed through high bandwidth memory at multiple terabytes per second. Each agentic task is dozens of those steps. Each query consumes single digit watt hours of electricity and a small but real amount of cooling water. Multiplied across a billion queries a day, the bill is in megawatts, megalitres, and tens of thousands of accelerator nodes.

Heinrich runs on a consumer workstation. CPU, not GPU. Standard DDR5, not HBM3. The per inference work is a coordinate lookup against a numpy float64 array, not a forward pass through a learned model. The hardware bill of materials for a personal Heinrich workstation is under five thousand Canadian dollars. The first physical product fits under fifty megabytes.

This is not an optimisation gain. It is the consequence of doing a different kind of physical work for every answer.

What you can audit

An LLM's reasoning is opaque. You can ask the model to "show its work," and it will produce a chain of thought string that looks like reasoning. But that string is itself a generated continuation, not a record of the model's actual computation. The real computation happened in the weights, and the weights do not explain themselves.

Heinrich's reasoning is a record. Every answer carries the activation cloud that produced it. The seed concepts, the propagation pathway, the amplitudes at each step, the source claim for every clause in the final sentence. You can ask why Heinrich said what it said and get a physical answer. These concepts activated, via these relationships, at these confidence levels, sourced from these specific claims in the Wikidata or ConceptNet index.

The audit is not a feature. It is the trace the system already kept to produce the answer.

Where this leaves the LLM

None of this means LLMs are going away. They are extraordinary at generative tasks. Long form writing, code synthesis, fluent conversation, creative composition. The work they do is real, valuable, and unmatched in those domains.

But the domain they are most often used for, including answering factual questions, providing trustworthy reasoning, and serving as the knowledge layer in agentic systems, is not the domain they were built for. They do it because nothing else of comparable capability has existed. The hallucination problem, the retraining problem, the audit problem, the energy problem, the dependency problem. Those are all consequences of using a generative architecture as a knowledge architecture.

Heinrich is the knowledge architecture. The plausible long term shape is a hybrid. Heinrich grounds the facts. A small generative model handles the prose. The frontier LLM, freed from carrying the world's structured knowledge in its weights, becomes a smaller, faster, more focused tool. The knowledge lives somewhere it can be addressed, edited, and inspected.

That is not a war between architectures. It is the right tool for the right work.

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