The problem
Language models optimize for the average — not for reality.
Science answers to the primary source — reality. A language model answers to text. Three fallacies follow, and they share one root: a statistical property of text is substituted for the thing itself.
01The reasoning fallacy. Most likely means consensus — ad populum, run as an architecture. A model trained before Copernicus would insist, fluently, that the sun circles the earth. Text buys correlation, never cause and effect — and from inside a system referencing only itself, insight and hallucination look identical.
02The relevance fallacy. Nearest is not relevant — and nobody speaks in chunks. People name things, and a name points at something persistent, with a history and relationships already attached. Retrieval by resemblance has no name to resolve, so it ships the haystack and trusts the model to find the needle.
03The human fallacy. People do not run on likely words. They run on affect — and the corpus assumes writing reflects what people feel, when most of what anyone feels is never written down. The signal that actually governs behavior is the one least present in the data.
So it flatters instead of resolving, and re-derives the same reasoning billions of times a day, discarding it when each conversation ends. Orealis corrects all three — a deterministic control plane for long-horizon agents: error measurable, memory auditable, help judged by what actually changed.
01The correction — causal structure
Causal reasoning from structure.
The primitive is the function, not the token — the finite cause-and-effect moves by which people satisfy needs. Life’s diversity looks infinite; at the level of function it reduces to the same closed set, run again and again.
Two layers: the lattice — that fixed map, identical for everyone — and the topology — the user’s life, embedded in it as coordinates. The model is a codec at the boundary — language in, coordinates out — and deterministic walks do the reasoning. Memory becomes placement, not retrieval. Reasoning becomes a path you can inspect, not a guess you have to trust.
The lattice is a control — the fixed reference that makes error measurable and keeps it from compounding. Core reasoning persists at its coordinates: derived once, walked cheaply forever. What it cannot answer names its own gap — researched, filed as a candidate, ratified by outcomes and, where stakes are high, by experts. Outcomes land at shared addresses and compound into population-level cause and effect. The users train the system. What accumulates is a causal graph of human reality.
Which is also the answer to scale. The lattice is not hand-authored — it is AI-assisted by construction: the model proposes, the structure verifies, people ratify. And it is recursive, so it grows inward rather than outward — new precision comes from composing coordinates that already exist, not from minting new ones. Set against retraining a trillion parameters to absorb one new thing, that is a different cost class entirely.
02The correction — typed objects
Names, not chunks.
Every person, group, and organization in the user’s life is a persistent typed object. A name resolves in real time to the right one, carrying its history and its edges to everything else — so disambiguation stops being a guess, and what a conversation is about is a fact the system holds, not an inference it re-makes each turn.
Context follows from that. What loads is what the conversation actually reaches — the needle, not the haystack — and it descends on demand, from summary to detail to the verbatim turn, so cost tracks the question rather than the length of the transcript.
This layer stands on its own. The typing earns its keep before the lattice does: resolution, provenance, and adaptive context are worth having whether or not anything is built above them.
03The correction — the human model
Averaged words don’t govern behavior. Affect does.
Affect governs human behavior — feeling is the signal that orients attention, decision, and action. So affect is the index: what a person feels marks what matters, and the system knows not just what is true but what is worth computing.
That layer is a computational model of human behavior built on one premise — the function of affect is to orient adaptive behavior. Orealis calls it adaptive affective metabolism. It aligns with the empirical evidence from neurobehavioral science — appraisal theory, somatic markers, allostasis — but it computes: one integrated causal runtime, not a shelf of theories that don’t.
Like the others, this layer is separable by design. Take the causal architecture without it and you still have a typed, deterministic reasoning substrate. Together, the three reason causally about what actually matters to a person.
Together
What the three produce.
- Alignment you can measureEvery suggestion carries an expectation — condition, outcome, horizon — audited against real progress, not against how the conversation felt. Predict, then check.
- Agency safety by constructionDeterministic gating cannot propose a route a person is structurally blocked from taking. No dead-on-arrival options, no false hope.
- Absolute provenanceEvery assertion cites the exact turn that grounded it. Contest one inference and that one coordinate is invalidated — cleanly.
- Compounding evidenceOutcomes land at shared addresses, so use accumulates into population-level cause and effect instead of evaporating with each conversation.
What this is — and isn’t
AI wrapper Core architecture.
Orealis, in one sentence: a fixed, shared addressing scheme for human life, with deterministic derivation over it — and a language model used as a codec at the boundary, not as the reasoner. That division of labor is the product.
- Not RAG, not vector memoryNothing is ranked by similarity — no embeddings, no top-k. Language resolves to a coordinate; walks do the rest.
- Not a per-user knowledge graphThe map is fixed and shared; only the user’s placements vary. That is what makes error measurable and memory comparable.
- Not fine-tuningNothing is trained. Knowledge lives in an editable catalog with the inference path visible — correctable in minutes.
- Not an agent frameworkFrameworks are plumbing for calling a model. This constrains what the model is asked to decide.
Symbolic systems failed at the language interface; language models fail at ground. Each solves the other’s problem — the pairing is the product.
The demonstration
“Hi, I’m Ori.”
The architecture runs live in a working companion — the clips alongside are it. Orí’s memory is a typed, causal model of what a person cares about: a structure it computes over, not a transcript it searches.
It maps their concerns, what feeds them, what blocks them, and which routes are genuinely open. It distinguishes feeling better from getting better — and audits whether its own help moved anything.
The name is the first syllable of Orealis — from Borealis, light appearing in darkness — and a Hebrew name meaning my light.