The engine is the truth, the model is the voice

Against the current

A researcher searches. I chose to search against the current.

The artificial intelligence of this decade moves in one direction only: bigger, and further from the ground. Models that double in size from one year to the next, trained on continents of data, housed in compute halls no one ever visits, and trusted more and more with the task of saying what is true. I took the other road, and this text explains why.

Searching against the grain, here, means aiming at four things the race to scale has pushed into the background. Frugality first: a system that runs on an ordinary machine and computes little, in place of a bottomless appetite for more. Sovereignty next: the data and the engine stay with whoever owns them, they do not travel elsewhere to be judged. Privacy, which follows from that and stops being a promise to become a property of the design. And above all, the most misunderstood of the four: giving the language model back the one job it truly does well, language and voice, by taking away the one we wrongly hand it, saying what is true.

The thread that ties these four aims together runs back to the neuroscience of memory, not to computer science. The idea holds in an image I stand by: take what a living brain does with a memory trace and transpose it into an algorithm of cold determinism, then add the language model only on top, as the warm voice of the tool. The cold holds the truth; the warm puts it into words. Everything that follows comes from that separation.

The observation

We gave machines speech before we gave them truth.

A large language model phrases things admirably and knows nothing. Give it a precise fact, a number, a date, a clause, a decision made eight months ago, and it returns the most plausible run of words, with no way of knowing whether that run is true. On public data, this passes most of the time. On yours, a client file, a regulated number, a contractual commitment, a medical result, the same plausibility turns against you: it takes on the look of an answer, and it lies.

The industry answers this flaw with RAG: retrieve nearby passages and hand them to the model. Retrieving neighboring text is not the same as remembering, and nearness has never made anything true. A setup like this ignores what actually makes a memory: contradiction, the decision that cancels an earlier one, the corroborated fact you tell apart from the fact seen only once, the matter that closes, the value a computation can check. It relocates the problem without solving it. The model no longer hallucinates about its memory; now it hallucinates about the context it was handed. The judge is still the model.

This is the original sin of AI systems laid over enterprise data: the model sits at the center. It is the one that decides what is true, at the instant you read it. As long as that seat does not move, no prompt, no fine-tuning, no guardrail makes the system reliable about a fact. You do not fix an architecture with an instruction.

The thesis

This shift goes deeper than an optimization: it changes the hand that decides. Truth is computed ahead of time, by code you can audit, test, and replay. When a question comes in, the model receives a fact already established, with its confidence and its source, and it has one thing to do, say it well. It settles nothing at the moment the user asks.

Here the metaphor turns technical. The deterministic engine is the cold part of the system: a computed, verified trace that does not change its mind between two readings, much like the engram that neuroscience describes as the physical substrate of a memory. The model is the warm part: speech, supple, sensitive to context, giving the fact a human form. The cold guarantees; the warm welcomes. Separating the two is a way of refusing to let speech pass itself off as memory.

From this reversal comes a discipline. It holds in seven principles, and these principles are technology-agnostic: they hold for an ERP, a CRM, an internal line-of-business tool as much as for a personal assistant.

The seven principles

1. 1

   Truth is computed, not generated.

   Facts are established at the point of entry, in plain and structured form, then read back as they are, with nothing reinvented on the fly. When a model helps extract them, it may not keep a single one it cannot point to in the source document; whatever it cannot show, it drops. A guessed fact is never served.

2. 2

   An identifier is proven before it is believed.

   A national ID, an IBAN, a case reference counts for what it claims to be only once it has passed the computation that validates it. An act number wrongly taken for a national ID fails that check, and so it is never displayed as one. Where the model would guess, the computation decides.

3. 3

   Every fact carries its source, its state, and its confidence.

   Nothing is shown bare. Every fact shows where it comes from, where it stands in its life cycle, proposed, in force, superseded, closed, and how far it can be trusted. The machine states its confidence, and the model is not allowed to inflate it. The brain already keeps this ledger: it holds the content of a memory side by side with its provenance, and it is when the thread of that provenance is lost that it starts to fabricate false memories (Johnson, Hashtroudi and Lindsay, 1993). Knowing whom a fact belongs to is a matter of honesty.

4. 4

   Contradictions are settled by computation, away from reading time.

   When two facts contradict each other, the resolution is computed, not negotiated in the moment. If a subtle case calls for a model's opinion, the model gives it off to the side, at write time, and its verdict is filed away as one more piece of data. At reading time, no model judges anymore: what gets served is a verdict already reached.

5. 5

   Memory has structure.

   Remembering is far more than laying hands again on neighboring text. It is holding together what belongs together, keeping vivid what matters and letting the rest fade without erasing it entirely, treating a contradiction or a decision as a first-class object. Living brains showed the way: what fires together ends up wired together (Hebb, 1949), and what we do not revisit fades along a curve we have known for more than a century (Ebbinghaus, 1885). The aim is a memory, where an index settles for retrieval.

6. 6

   The trust boundary lives in the code.

   Content of doubtful origin, an email, a web page, anything at all that anyone can fabricate, is marked as such and can trigger no action. The slightest effect on the world, writing, sending, modifying, goes through a human confirmation and leaves a trace. The rule allows no exception: you never place within the model's reach a piece of data or a power it is not allowed to emit. The ban lives in the code, not in the prompt.

7. 7

   Confidence governs behavior.

   "I don't know" is a full and legitimate answer. The psychology of eyewitness testimony showed this half a century ago: you can be sure of yourself, precise down to the smallest detail, and wrong from start to finish (Loftus and Palmer, 1974). A confident answer tells you nothing about whether it is right. When certainty is missing, the value is not even handed to the model: abstention is not hoped for from an instruction, it is made unavoidable by what the code chooses to withhold. A trustworthy system knows how to stay silent, and it would rather decline than mislead.

What an organization gains

This design was built for the places where a mistake has a cost, not for a demo. On compliance, it offers plainly what the GDPR, the AI Act, and regulated professions ask for: a source on every assertion, auditable actions, an abstention that leaves a trace. A system you can certify, where the model-at-the-center stays, at best, probable.

Then comes ordinary trust, the kind of everyday use. The machine asserts what it has computed and shows it; when it doubts, it says so; when it does not know, it stays quiet; and it never hides where it got what it puts forward. That is what makes it usable where plausible-but-wrong is expensive.

The cold has a price too, in the good sense of the word. At reading time, a factual answer is a computation, not a call to the model: faster, cheaper, and above all reproducible, the same question giving the same answer. Inference goes back to what it should have stayed, a commodity you draw on only at the edges. This is frugality held by construction, not promised in a slogan.

Sovereignty remains, which was the starting point. The engine of truth lives with you. The model, for its part, is interchangeable and may never leave the machine. What gains value over time, the fabric of facts, decisions, and contradictions, belongs to you and thickens as you use it.

The existence proof

This thesis did not stay on paper. It is implemented, deployed, and audited in Hygur, a local-first digital twin that takes in a person's informational life, their email, their calendar, their documents, and answers about their facts without ever inventing.

The demonstration takes a minute. Ask Hygur for a precise number: it returns the value, sourced, together with its confidence, or it honestly declines when the attribution stays ambiguous. Then cut off the language model. Hygur goes on searching, retrieving, answering about its facts. The truth had not moved, because it did not depend on the model. The engine held the truth; the model was only the voice we had just silenced.

The core of Hygur already bears the name: the Engram. In neuroscience, the engram is the physical trace of a memory, the substrate where a memory is inscribed and later found again. The word is Richard Semon's, who coined it in 1904 to name the lasting mark a stimulus leaves in a living organism. A century later, Susumu Tonegawa's team made it observable, then manipulable: by switching the neurons of a memory back on in a mouse, they brought the memory itself back (Liu, Ramirez et al., Nature, 2012). The engram stopped being a vague picture and became a trace you go looking for and find again. That is precisely what we set against the model's hallucination: a verified, sourced, computed trace. Recall, not hallucinate: the recall of a fact on record rather than the invention of a plausible one.

Hygur is the reference implementation, the proof that the thesis stands up away from the whiteboard. The method travels. It is the method this text defends, and the method an organization can carry into its own tool, its ERP, its CRM, its ITSM, its business database, without having to recode Hygur.

The call

If you have deployed a language model on your data, you already know the wall: it is brilliant, and it lies about your facts. The gain will come from a better division of roles, not from a bigger model. Give the truth back to a deterministic engine. Keep the model for what it does well, speaking.

Build the engine. Give it the truth, the provenance, the confidence. Let the model be the voice, and never the judge. It asks more than a coat of API paint, and that is exactly what makes it hold: a system you can audit and certify, and one that lasts.

The engine is the truth. The model is the voice. Recall, not hallucinate.