Research
One root cause. One lever.
Every wall frontier language models hit in 2026 — runaway compute, catastrophic forgetting, hallucination, opacity, no clean correction — is a symptom of a single design decision: knowledge, language, and reasoning entangled in one weight space. Resonance removes the precondition instead of managing the cost. It contains no pretrained model at all. This page lays out the thesis, what has been built and measured so far, the open questions that decide the company, and where we believe the architecture can go.
Everything is entangled in one weight space.
A conventional model bakes the sum of human knowledge, every language, and its reasoning patterns into one shared parameter space, learned together as a side-effect of predicting the next token. It does not look anything up. Every fact is reconstructed by running the whole space forward, because there is no separate store to consult — only weights that re-derive the answer each time.
Three consequences follow directly, and each has become a wall the field is now spending against.
Capacity is the knowledge ceiling
Because knowledge lives in the parameters, what a model can know is bounded by how big it is. Knowing more means being bigger, which means retraining.
Updating means overwriting
Because everything shares one space, learning something new perturbs the weights that held something old — and a frozen model can update exactly zero facts mid-deployment.
Every query pays for the whole store
Because answers are reconstructed rather than looked up, a trivial query draws on the same machinery as a hard one. Compute is coupled to model size, not task difficulty.
The industry's two moves — scale the compute, or scope the model — both keep the entanglement and try to manage its cost. Neither questions the monolith.
— Ioma Labs, design philosophyRemove the precondition.
Resonance refuses the entanglement at the start rather than unwinding it later. There is no pretrained model inside and no training corpus behind it. Knowledge, relations between knowledge, and language are each held in an explicit, gradeable store — and one mechanism operates on all three: a unit of knowledge is stored, retrieved, scored against feedback, and reinforced or weakened over time.
The system boots completely empty. The first message it receives becomes its seed; from there, every capability is acquired incrementally from graded human — or expert-approved automated — feedback. Four properties define the shape:
- Knowledge as explicit chunks with strength scores, ranked at retrieval by similarity × strength — stored, never reconstructed, and readable off disk by a human.
- Relations as independent records — typed edges between facts, where confirming or rejecting one cannot disturb another. Disentanglement is structural, not a training trick.
- Language that is earned — a learned lexicon and a rule-first grammar, taught through the diff between what the system said and the tutor's correction. It can only speak words it has been taught.
- Feedback as the native input, with every training event recorded permanently, with tutor attribution, to an append-only ledger — so the entire learning history is replayable, including onto a larger future encoder.
Once knowledge, language, and reasoning are physically separate, the symptoms decouple. Hallucination loses the mechanism that fabricates. Correction becomes one graded event instead of a retrain. Answering becomes a lookup, not a generation-scale computation — which is why the whole system runs on ordinary CPU hardware.
We are not building a cheaper version of the current thing. We are building a different cost structure for intelligence.
— Ioma Labs, design philosophyWhat is built, and what we've measured.
The architecture is complete: all three layers implemented, verified, and deployed as a live service with crash-safe writes and versioned off-box backups with a tested restore. In July 2026 the full build passed a hardening audit — every document and code file read, every suspected defect reproduced before being reported, every fix measured before it shipped. Four critical defects were found and fixed. These are the numbers that survived.
Old-fact retrieval margin under forced interference, versus +0.012 without replay. Teaching more no longer means forgetting — measured, not claimed.
Signal-based sleep consolidation replaced the O(N²) centrality model. The gap widens with scale; consolidation is no longer a scaling concern.
Sub-millisecond to ~20k. Four approximate methods were tested and rejected on accuracy — the real ceiling is encoder generalization, not search speed.
A committed teach survives a hard kill via the write-ahead log; a full restore onto a new host is verified, encoder geometry included.
The scaling investigation reached one throughline: retrieval speed, dimensionality, and index tricks are all downstream of whether the encoder places never-seen text near the right taught neighbors. That property — generalization — is emergent from training volume, and it is watched continuously by a built-in held-out probe, live now, that queries taught facts with paraphrases the system was never shown. It is the central scaling metric from the first day of teaching.
Underneath everything sits the durability guarantee we build to: the source text of every fact, the current learned state, and the complete replayable training history are preserved independently. Any future change of encoder size or storage backend is a migration, never a loss. No scaling decision can strand the knowledge.
The questions that decide the company.
We bring a measurement plan, not a claim. Each open question below has a test attached, and each starts running on day one of teaching. None of them determines whether the road exists — the durability design guarantees every fallback is a migration — but together they set its price and its timeline.
| Question | Why it's the one that matters | How we measure it |
|---|---|---|
| Sample efficiency | The binding question: corrections-per-unit-knowledge is the exchange rate that prices the entire journey to breadth | Tracked continuously via the held-out generalization probe from the first taught fact |
| Encoder generalization | Does unseen phrasing land near the right taught facts? This single signal gates fast search, the inference engine, and large-scale operation | Held-out paraphrase separation over training time; rising = generalizing, flat = the trigger to grow the encoder by ledger replay |
| Catastrophic interference | Does new teaching erode old discriminations in the shared encoder? Rehearsal replay neutralized it at small scale; scale must confirm | Forced-interference retention tests, repeated as the mesh grows |
| Parallel multi-tutor teaching | Can many experts teach disjoint domains simultaneously without cross-talk? The board-of-experts model depends on it | Per-tutor attribution is recorded on every event, so cross-domain interference is measurable — and correctable by replay |
| The value floor | Correctability cuts both ways: one grade can install a falsehood as firmly as a fact. A taught system has no inherited floor under what it can be taught | Corroboration requirements, tutor trust gating, and abuse controls on the teaching endpoint — funded in this round, before open teaching, not after |
A standing principle, carried from the build: measure before re-architecting. This project has twice mistaken an undertraining problem for an architectural one. Apparent limits are treated as training artifacts until an experiment shows otherwise.
— Ioma Labs, research agendaWhere we believe it can go.
Can this reach frontier-benchmark capability? The honest answer: the audit of the full codebase found no raw impossibility — nothing this architecture cannot be made to do in effect, given effort and resources. What remains is the bill. Here is its real shape.
- Knowledge breadth — the biggest line item. Frontier models bought their breadth with the entire internet; Resonance buys its breadth with graded corrections. Domain experts set each field's skeleton; expert-approved automated pipelines deliver exhaustive nuance at a rate no human fleet could. The constraint is throughput, not capacity — precisely what funding converts into speed.
- Encoder generalization — earned through the same teaching. If the current encoder proves too small, growth is a warm replay of the permanent training ledger onto a larger one — a rebuild that loses nothing, by design. Generalization also unlocks fast approximate search for free.
- Fluency and style — same loop, different layer. Frontier-grade language is taught through the identical mechanism as knowledge. Early output is plain by design; plain is a stage, not a ceiling.
- Reasoning and creative synthesis — trained and switched on. The inference engine — a relation proposer that synthesizes candidate concepts present in no single taught fact — is built and deliberately gated until the encoder matures. At a large confirmed mesh, recombination across domains functions as generative capacity, with provenance. Every novel idea surfaces as a checkable proposal built from graded material.
- Tools and affect — the designed layers. Tool use slots in as a second exit from retrieval; affect as a taught character-and-mood layer. Both are designed with mechanics worked out, neither yet live.
What "equivalence" honestly means here is not identical behavior — it is equivalent effect, reached a different way. On taught territory, the aim is to match or beat frontier models on the axes that matter for knowledge work — accuracy, provenance, correctability — while approximating their breadth and fluency through volume of teaching. Benchmarks built for pretrained models measure the absence of a corpus, not the quality of this architecture. The honest instrument is the learning curve: capability as a function of corrections delivered, read against frontier performance on the same tasks.
What we'll publish.
We'd rather show than claim. As teaching begins, the numbers publish as they accrue — including where the approach doesn't win yet.
The learning curve
Capability as a function of corrections delivered, on the first taught vertical — read against frontier performance on the same tasks. The round's central deliverable.
Publishes with teachingHeld-out generalization
Does phrasing the system was never shown land near the right taught facts? The probe runs continuously; its trace publishes with it.
Publishes with teachingThe hardening record
The July 2026 stress-test audit — defects, fixes, and measurements — prepared as a public methodology note on verifying a from-scratch learning system.
In preparationWorking on taught systems, or want to test ours?
The hardest questions here are open and measurable. If this is the problem you want to spend your time on — or you want to back the work — we'd like to hear from you.