Resonance
It starts empty. That's the point.
Resonance is a from-scratch, feedback-trained memory and language system. It contains no pretrained model and consumes no training corpus. It boots blank; the first message it receives becomes its seed, and every capability it will ever have is acquired incrementally from graded feedback. What it is taught is what it holds — exactly, indefinitely, with receipts.
One mechanism, three layers.
A unit of knowledge is stored, retrieved, scored against feedback, and reinforced or weakened over time. Resonance applies that single mechanism to facts, to the relations between facts, and to language itself — and every turn runs all three grading channels in strict order, lowest layer first.
Retrieval
A byte-level encoder maps text to meaning vectors. Facts carry individual strength scores, and a rehearsal-replay mechanism re-anchors past learning alongside every new correction — so new teaching reinforces rather than overwrites old knowledge.
Relations
Typed edges between facts — supports, contradicts, precedes, implies — stored as independent records. This is what lets the system reason across facts, and eventually beyond them, rather than just recite them.
Language
A learned lexicon, a rule-first grammar — a correction installs a broad rule immediately; later contradictions carve bounded exceptions — and a word-order engine. Vocabulary is earned, never imported.
There is no confidence gate and no "I don't know." Once seeded, Resonance always answers with its best attempt — because a weak answer is correctable, while a refusal teaches nothing.
— Ioma Labs, design principleWhat the architecture guarantees.
These are not features that were added. They are consequences of how the system is built.
No hallucination of the untaught
It can only output what has been grounded. Outside its taught domain, the worst case is a plain or weak answer that invites correction — never a confident fabrication.
Surgical correctability
A wrong fact is fixed by one graded correction, in seconds, permanently — not a retraining cycle that puts every other capability at risk.
Provenance by construction
Every answer traces to specific stored facts and confirmed relations. The answer literally cannot exist without its sources.
Epistemic transparency
Every belief carries a strength, a hit count, a confirmation flag, a timestamp, and a tutor attribution. The entire belief state can be opened and read by a human.
Sycophancy immunity
Belief strength changes only through explicit graded feedback — never through tone, insistence, or social pressure. You cannot argue it into agreeing; you can only teach it.
Zero-loss durability
A committed teach survives a hard crash. Source text, learned state, and the complete training history are preserved independently — no scaling decision can strand what it knows.
Two more follow from the economics: answering is a lookup, not a generation-scale computation, so it runs on ordinary CPU hardware — and self-hosted means the instance an organization teaches is theirs, entirely, down to air-gapped deployment.
One property, both directions.
Resonance is an exact mirror of its curriculum. It holds precisely what it is taught, at full strength, indefinitely — it does not average its sources into a blur. With a world-class curriculum, that is the product: a specialist that does not diverge from its experts. Honesty requires naming the other direction too.
Trust is concentrated
A pretrained model dilutes a bad source across a corpus; Resonance concentrates trust in its tutors. Corroboration requirements, tutor trust gating, and per-tutor attribution are the funded remediations — but the dependence is the design.
Recovery is replay, not edit
A bad training event can't be surgically excised from shared encoder weights. The designed recovery — replay the attributed ledger onto a fresh encoder, omitting the bad events — is lossless, but it is a rebuild, not an undo.
The ramp is ours, not yours
Every new domain starts empty and is taught to depth before it ships. The buyer receives a fully-taught specialist; the teaching bill is paid internally.
High-trust knowledge work.
Anywhere a wrong answer is expensive and "show me why you believe that" is a requirement, not a preference. These buyers aren't underserved by today's models — they're disqualified from using them, because they cannot audit a belief or fix an error.
Regulated and safety-critical domains
Compliance and policy teams, legal and regulatory reference, clinical knowledge, financial operations, safety-critical engineering documentation. Resonance's native properties are their entry requirements.
Institutional memory
The org whose answers live in a thousand documents and twelve veterans' heads. Taught once by the people who actually know, it becomes an auditable memory that never leaves the building.
Air-gapped deployments
Government, defense, healthcare — any buyer for whom "the model runs on our hardware and contains only what we put in it" is the first question. No frontier-model vendor can offer this shape at all.
Who it is not for, today
Anyone needing open-domain, zero-setup breadth on day one, or open-ended creative generation. That buyer is served by the incumbents — at the incumbents' margins, with the incumbents' failure modes.
Where it stands today.
The build is done and live. The next eighteen months are about teaching it.
-
Built — Complete
Architecture implemented end to end
All three layers built and verified, wired to a live deployed service — with crash-safe writes, a permanent attributed training ledger, and versioned off-box backups with a tested restore.
-
Hardened — July 2026
Full-build stress test passed
Every document and code file audited; four critical defects found, fixed, and measured. Retention under new learning verified with rehearsal replay. The audit is written down, not folklore.
-
Now — First domain
Stand up the board of experts
Domain specialists set one vertical's conceptual skeleton; expert-approved automated curriculum fills nuance at volume. The learning curve and held-out generalization numbers publish as they accrue.
-
Then — Scale the loop
Parallel teaching, measured
Validate multi-tutor teaching without cross-domain interference, on owned training hardware. Every event is attributed, so interference is measurable — and correctable by replay.
-
Later — Commercial launch
Fully-taught specialists, and personalities
A generous free tier and a $10 paid tier, on self-hosted economics. Taught personalities ship on top of the architecture — Calliope, the name our earlier prototype carried, returns as the first of them.
Where Calliope went.
Calliope was our earlier prototype — the project that taught us what this architecture had to become. Resonance is its final form: the framing changed from a distributed arrangement of pretrained parts to a system with no pretrained model inside at all. The name doesn't retire. In the commercial release, Calliope returns as a personality of Resonance — a taught character and voice on top of the architecture, not a separate system.
Two instances taught by different tutors know the same field but think and speak differently. A personality, here, is not a system prompt — it is taught, and it is earned the same way everything else is.
— Ioma Labs, internal framingWant to see the mesh, or help teach it?
We work with a small circle of partners, investors, and domain experts. If you'd like a demonstration, or a seat on the board of experts, reach out.