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.
FIG. 1 — One Resonance teaching turn, schematic. Read bottom to top. DASHED LOOP — graded feedback returning to every layer
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 philosophyOne 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.
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.
Relations — how facts connect
Typed edges between facts — supports, contradicts, implies — stored as independent records. Confirming or rejecting one relation cannot disturb another.
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.
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.
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.
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.
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.
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.