Resonance

A different kind of intelligence.

Ioma Labs is an AI research company. Our system, Resonance, is a from-scratch, feedback-trained memory and language system — no pretrained model inside, no training corpus consumed. It boots empty; the first message it receives becomes its seed, and everything it will ever know is acquired one graded correction at a time. The result is a system whose every belief is grounded, auditable, and correctable — properties of the structure, not features bolted on.

Resonance teaching-turn schematic A schematic of one Resonance teaching turn, rising from a mesh of stored facts, through confirmed relations and the learned language layer, to a spoken answer — with graded feedback looping back down into every layer. Hover or focus a stage to read it. FLOW 01 KNOWLEDGE MESH 02 RELATION MESH LEXICON · GRAMMAR · ORDER 03 LANGUAGE LAYER 04 SPOKEN ANSWER 05 — GRADED FEEDBACK

FIG. 1 — One Resonance teaching turn, schematic. Read bottom to top. DASHED LOOP — graded feedback returning to every layer

01 — Problem

Almost every problem is the same problem.

Every wall frontier models hit in 2026 — runaway compute, catastrophic forgetting, hallucination, opacity, no clean correction — traces to one design decision. Knowledge, language, and reasoning are entangled in a single weight space, so every answer is reconstructed from scratch rather than looked up. Resonance removes the precondition instead of managing the cost: it contains no pretrained model at all. Knowledge lives in an explicit, gradeable store outside any weights.

Separate what a system knows from how it says it, and the symptoms decouple: auditability, correctability, and hallucination resistance stop being features you bolt on and become properties of the structure.

— Ioma Labs, design philosophy
02 — Architecture

One mechanism, three layers.

A unit of knowledge is stored, retrieved, scored against feedback, and reinforced or weakened over time. The same mechanism operates on facts, on the relations between them, and on language itself.

Layer 01

Retrieval — what's near

A byte-level encoder maps any text to a meaning vector. Facts are stored as chunks with individual strength scores; retrieval ranks by similarity × strength.

Layer 02

Relations — how facts connect

Typed edges between facts — supports, contradicts, implies — stored as independent records. Confirming or rejecting one relation cannot disturb another.

Layer 03

Language — how to say it

A learned lexicon, a rule-first grammar, and a word-order engine. The system can only speak words it has been taught — vocabulary is earned, never imported.

The loop

Feedback is the native input

Every turn runs three grading channels in strict order — retrieval, relations, then the spoken sentence. Twelve plain-language verbs carry every correction; the tutor never sees the plumbing.

The ledger

Every event, on the record

Each teach is crash-safe and recorded permanently, with tutor attribution, to an append-only ledger. The complete belief state on disk is human-readable.

Why it matters

Correct by construction

It cannot hallucinate what it was never taught, a wrong fact is fixed by one graded correction, and every answer traces to the specific facts and relations behind it.

03 — Product

Built, hardened, and live.

Resonance is not a thesis with a demo attached. The full system is deployed end to end — chat, feedback, versioned off-box backups with a verified restore — and passed a full hardening audit in July 2026. The build is done; the next chapter is teaching it, domain by domain, through a board of experts. In the commercial release, taught personalities sit on top of the architecture — Calliope, the name our earlier prototype carried, returns as the first of them.

04 — Research

We measure before we claim.

Resonance improves as a plotted line — capability as a function of corrections delivered, checked continuously by a built-in held-out generalization probe. As teaching begins we'll publish the learning curve, the open questions, and the places the approach doesn't win yet.

The architecture is built. This round teaches it.

We work with a small circle of partners and investors. If you want to help teach a different kind of mind — or back the work — we'd like to hear from you.