Most tools handle brain data by learning statistical patterns: this pattern usually means attention, that one usually means fatigue. That works until it matters. A pattern-matcher can flag that something looks different today — it cannot say what changed, whether it means anything for this person, or what would help.
Our answer has two halves that only work together: the Quantum Virtual Mind (QVM) — a quantum–classical model of the dynamics themselves — and a layer of specialist AI agents wrapped around it. The QVM supplies a state you can interrogate; the agents do the work of getting data to it and turning its answers into something actionable. Neither is the product on its own.
Live and recorded BCI streams from standard equipment. Rhythms and inter-region coupling extracted continuously, with each subject building their own history rather than matching a population template.
The core. Those dynamics reconstructed as an executable quantum-classical state model: how activity routes through the network, where it dissipates, how the whole settles. Built so physical consistency holds by construction.
Specialist agents work the whole lifecycle — accessing data, engineering it, developing and checking the model, running perturbation studies, and designing what to do next.
Superlearning reads the healthy direction toward capacity, restoration reads it back toward a settled baseline, and the disorder-studies programme reads the clinical one. Both are proposed applications designed to write an auditable state record.
A BCI recording is a narrow window: a few dozen channels of electrical activity summarising the collective behaviour of many millions of cells. The model captures what appears in that window — the rhythms, the couplings, and how they evolve — faithfully enough to be interrogated, and scoped precisely enough that every claim we make about it is testable.
Why the smaller claim is the stronger one
A model of the biological brain would be unfalsifiable at present. A model of BCI dynamics can be tested against held-out recordings, compared against conventional methods, and shown to be wrong. We prefer the version that can be checked.
Real systems cannot spontaneously gain energy; they settle rather than explode. A model merely fitted to data can violate that and still look plausible. Ours is built so those rules hold automatically — structural, not a penalty that can be overwhelmed, and enforced at every stage of training.
Dynamics are described as named parts — the balance between excitation and inhibition, delays on long-range connections, the strength of particular hubs. When something changes, the model says which part changed: a researcher can interpret that, and a reviewer can audit it.
Not a recommendation engine bolted onto a finished model — specialists operating at every stage.
Finding recordings, checking consent and provenance, and tracking which subject and session every file belongs to. Unglamorous, and the stage where most downstream errors originate.
Cleaning, aligning, and conditioning signal so a model sees comparable input across sessions. Each subject accumulates their own history rather than being matched to a population template.
Fitting the state model to a person, then checking it — against held-out data, against physical constraints, and against the simpler baseline that might explain the result just as well.
Measuring how far a state has moved from that person’s baseline, in which component, and in which direction. The same operation serves the learning, restoration, and clinical questions.
Turning a measured displacement into a proposal: a learning track, or a candidate intervention. Presented as an argument a person can interrogate, with its reasoning attached.
Each owns one part of the lifecycle and each can be inspected on its own. Six work every subject; two are direction-specific — one for learning, one for clinical research.
Tracks where every recording came from, what consent covers it, and which studies may use it. Refuses a request that would pull data outside its permitted scope, and keeps the provenance chain attached to the model so any later result can be traced back to its source.
Harmonises recordings made on different equipment, at different sampling rates, with different electrode layouts. Flags sessions too noisy to use rather than quietly passing them through, so a poor recording never becomes a confident-looking result.
Fits the state model to a subject’s recordings and reports fit quality honestly — including when the data will not support a model at all. A refusal here is a feature: a twin fitted to insufficient data is worse than no twin.
Probes how the model responds when something changes — a learning session, a medication, the passage of months. Quantifies the departure from that person’s own baseline and identifies which components moved.
Reads the healthy direction. Turns a learner’s current position into a sequence: what to study next, at what difficulty, and when to stop for the day. Every recommendation carries the evidence that produced it.
Reads the clinical direction. Ranks candidate approaches for a researcher to evaluate, each with its reasoning and its uncertainty stated. Proposes; it never prescribes.
Cross-examines the other agents. Where two reach different conclusions from the same model, it surfaces the disagreement rather than picking a winner — a split verdict is information, not a fault to be hidden.
Writes the audit trail: what was tried, what moved, what was advised, and on what evidence. This is what makes the output reviewable by a clinician, an ethics board, or a regulator months later.
The agents advise. People decide.
Nothing in this layer acts autonomously on a person. Its output is an argument presented to a researcher, educator, or clinician — with the evidence attached and the dissenting view included. Because every agent reads the same physically consistent model, none can build an argument on an impossible brain state.
The three applications are proposals, and by design they are not separate products sharing a codebase.
| Agent capability | Superlearning — healthy | Restoration — recovery | Clinical — disorder studies |
|---|---|---|---|
| Data access | Learner sessions, consented through the platform they already use. | Everyday sessions on consumer hardware, opted into by the person themselves. | Longitudinal clinical recordings, under protocol and ethical approval. |
| Data engineering | Short, frequent sessions on consumer hardware. | Short recordings in unstructured settings — at home, at the end of a day. | Sparse visits recorded years apart on different clinical systems. |
| Model development | A twin per learner, refitted as they progress. | A twin at the person’s settled baseline, held as the target to return to. | A twin at the person’s healthy baseline, held as reference. |
| Perturbation study | Which changes move this learner forward, and how far they have moved. | Which practices actually move this person toward that baseline, and how long it holds. | Which changes characterise this condition, and how far it has moved them. |
| Design & recommendation | The next thing to study, and when. | Which session to run now, and how long to let it play. | Candidate approaches worth investigating, ranked with reasoning. |
| Record keeping | A learning record: what was tried, what moved, what was advised. | A record of what was practised and what settled, kept by the person. | The same trail, auditable by a clinical researcher. |
Read the perturbation row again
Learning and disorder are the same measurement — a departure from the person’s own baseline — distinguished only by direction and by whether you want to encourage it or reverse it. That is why one stack is designed to serve both.
Where the work stands
The state model is built and benchmarked in house, with manuscripts in preparation and wider validation underway. The agent layer is specified and being built. All three applications are proposals described here as designed, not as shipped. Nothing in this stack is a medical device.