SC‑WBD

What an individualised brain model makes possible

Almost everything worth doing to or with a nervous system needs a model of that nervous system. Here is what becomes reachable when you have one, in the order it becomes reachable, with the measurement that decides each step.

The average brain is not a brain. It is a statistic, and no intervention is delivered to a statistic — which is why the most valuable capabilities in neurotechnology are all blocked on the same missing object: a model of one particular person's brain, good enough that you can ask it a question you have not already measured the answer to.

That object is what this programme is building. The four sections below are ordered by how close each one is: the first is a problem of engineering and measurement, the last is a research programme. Each says what would have to be true, and what we would measure to find out.

Stimulation that knows where it landed

Transcranial magnetic stimulation is approved and reimbursed for treatment- resistant depression. Its weakness is not the physics — the induced electric field is computable, and we compute it, validated against analytic references. The weakness is that the loop is open. Targeting is done by scalp landmark or a group atlas, the response is assessed by symptom scales weeks later, and nothing in between tells you what the pulse did to the network.

An individualised forward model closes that loop: predict the network consequence of a candidate coil pose, measure the consequence, and use the discrepancy to improve the model of that specific person. The same argument runs below the clinical bar — attention, sleep, anxiety regulation — where the evidentiary standard differs but the need for individualisation does not.

What decides it. Our own measurement says the current state representation is not yet sufficient: for 98.1% of cortical locations, a parcel-level model's prediction of the sensor response to focal stimulation is wrong by more than half its own magnitude. That is a quantified obstacle with a known direction of fix — orientation, not spatial resolution — and closing it is the nearest concrete milestone on this page.

Interfaces that still work tomorrow

Two capabilities usually discussed together, and very different in difficulty.

Computer control from EEG partly exists today: a handful of discrete intentions, decoded reliably enough to drive an interface. The gap between a laboratory demonstration and daily use is almost entirely stability — across sessions, across days, across a person's own state. That is structurally a problem of modelling a person rather than a population, which is exactly the problem an individualised generative model is for. A decoder fitted to yesterday's data has to be refitted; a model of the person can be updated.

Language decoding from non-invasive recording is much harder, and the honest statement is that published work is far from continuous open-vocabulary decoding from EEG. The interesting question is whether decoding against a person-specific forward model is easier than decoding against a discriminative map fitted to their data. There is a real reason to think so — the forward model constrains the solution with anatomy and dynamics the discriminative map has to learn from scratch. We have not shown it, and showing it is a falsifiable experiment rather than an aspiration.

Non-invasive intervention with feedback

Transcranial focused ultrasound is the most interesting modality on the horizon: it reaches deep structures with spatial precision magnetic stimulation cannot approach, and its effects are reversible. A bidirectional loop — read the state, compute an intervention, deliver it, read the consequence — is the natural endpoint, and it is the point at which a forward model stops being a convenience and becomes the thing that makes the system possible at all.

It is also where the safety argument has to be strongest, and where this project has spent effort well ahead of capability. The intervention planner refuses by default: a plan declaring intent to drive real hardware or to be applied to a person is rejected unless a record exists that a preliminary review occurred with an approving outcome, and a valid authorisation record is necessary and explicitly not sufficient. The gate does not open because a date passed — there is a test that advances the clock to 2027 to prove it. Thermal dose limits are enforced on live plans specifically because two of those axes had no producer anywhere in the codebase, which meant a live plan could otherwise have been silently unchecked on the one thing that burns tissue.

Why the refusals come first

It is far easier to build a refusal into a system that cannot yet do anything than to retrofit one into a system that can. The gating above is out of proportion to what SC-WBD can do today, and that is deliberate.

Modelling what an experience does

If a generative model of an individual's dynamics is good enough to predict the consequence of a magnetic pulse, the same machinery in principle predicts the consequence of any input — a sentence, an image, a piece of music. The intervention stops being a coil and becomes content: individualised, non-invasive interventions delivered through the sensory channels a person already has.

The obvious objection is the right one: this is what every therapist, teacher and film director already attempts, with a model that lives in their head. The proposal is not that the effect is new. It is that the model could be explicit, individual, and improvable against measurement. In the owner's framing it is

the computational scaffolding dogs and humans and other mammals use for empathy

— which is worth taking as a technical statement rather than a poetic one. Predicting what an input will do to another mind is what social cognition does, continuously and cheaply, in species with far less machinery than we propose to build. A mammal running a fast, wrong, useful model of a conspecific is an existence proof that the problem is tractable at some fidelity. The research question is which fidelity suffices for which purpose.

Further out sits the framing that motivates the programme: an exocortex — cognitive capability that is genuinely yours, coupled tightly enough to feel like extension rather than tool use — and past that, a model of a particular person faithful enough to act as they would. The distance from here to there is not a matter of scaling. It is a sequence of unsolved problems, several of which we can now name precisely because we have tried and failed at the first one.

What would tell us we are wrong

A thesis without falsifiers is a pitch. Four results, any of which would substantially undercut this programme, and each of which we are positioned to measure:

  • Structured regional state does not beat a pooled vector at matched capacity. This is the pre-registered hypothesis of the current work. If the answer is no, the first differentiator is wrong and the architecture is more complexity than the evidence supports.
  • The anatomy prior does not resolve operator families. Partially true already — cytoarchitecture failed globally, and visual and somatomotor cortex do not separate on any evidence block we hold. If the binary unimodal/association split also fails to matter, then “region-indexed state” is a distinction without a difference at the resolution we can measure.
  • Individualisation does not beat a population model on any endpoint that matters. The entire argument rests on the average brain being the wrong object. That is an empirical claim and it is testable.
  • The forward model cannot predict a perturbation it has not seen. No checkpoint here has been trained on perturbational data, and the mapping from stimulation to response is flagged unvalidated in code for that reason. If predicting held-out perturbations turns out to require training on the perturbations themselves, the closed loop does not close.

Where the work actually stands

Nothing on this page has been demonstrated on a person, and no model published here yet beats its baselines. What exists is the machinery underneath: a typed schema that fails closed, a 414-region anatomy prior, six dynamics backends, a validated field solver, an identifiability laboratory, and two published checkpoints whose measured results are both negative and both public. The overview lists every one of them with its state, and the engineering notes describe, in detail, what went wrong building them and which run each lesson belongs to.

That is the honest position, and it is also the reason to take the plan seriously: the obstacles on this page are named, quantified and attached to experiments, rather than deferred to scale.