Where this leads

The case for building individualised whole-brain models is that almost every interesting thing you might want to do to or with a nervous system requires a model of that nervous system — not a model of the average one. The average brain is not a brain. It is a statistic, and no intervention is delivered to a statistic.

The near ground: intervention that knows where it landed

Transcranial magnetic stimulation is already used clinically for depression. Its central weakness is not the physics — the induced field is computable, and computing it is a solved problem we have implemented and validated against analytic references. The weakness is that nobody can say what the pulse did to the network. Targeting is done by scalp landmark or by a group atlas, response is assessed by symptom scales weeks later, and the loop between them is open.

What an individualised forward model offers is closure of 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 applies to wellness applications well short of clinical treatment — attention, sleep, anxiety regulation — where the bar for evidence is different but the need for individualisation is identical.

What would have to be true first. Our own measurements say the current state representation cannot answer this question: for 98.1% of cortical locations, a parcel-level model's prediction of the sensor response to a focal stimulation is wrong by more than half its own magnitude. That is a concrete, quantified obstacle with a known direction of fix — orientation, not resolution — and it sits between here and anything on this page.

Reading: language and control from EEG

Two capabilities that are often discussed together and are quite different in difficulty.

Computer control from EEG is the easier one and partially exists in the world today: a handful of reliable discrete intentions, decoded robustly enough to drive an interface. The gap between laboratory demonstrations and daily use is mostly about stability across sessions and across days — which is, structurally, a problem of modelling a person rather than a population, and therefore the problem an individualised model is for.

Language decoding from non-invasive recording is much harder, and honesty requires saying that the published state of the art is far from continuous open-vocabulary decoding from EEG. The interesting question is whether a generative model of an individual's cortical dynamics changes the problem — whether decoding against a person-specific forward model is easier than decoding against a discriminative map fitted to their data. We believe it is. We have not shown it.

Writing: bidirectional interfaces without surgery

Transcranial focused ultrasound is the most interesting non-invasive modality available, because it reaches deep structures with spatial precision that magnetic stimulation cannot, and because its effects are reversible. A bidirectional interface — read the state, compute an intervention, deliver it, read the consequence — is the natural endpoint.

It is also where the safety argument has to be strongest, and where this project has spent real effort ahead of any 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. A valid authorisation record is necessary and explicitly not sufficient. The gate does not open when a date passes — there is a test that advances the clock to 2027 to prove it. And 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.

The further ground: modelling what an experience does

Here the speculation becomes genuine speculation.

If you have a generative model of an individual's neural dynamics that 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 starts being content: individualised, non-invasive cognitive interventions delivered through the sensory channels a person already has.

That is a large claim and it deserves its obvious objection: 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, this is:

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

Which is the right way to describe it, and worth taking seriously as a technical statement rather than a poetic one. Predicting what an input will do to another mind is exactly what social cognition does, continuously and cheaply, in species with far less machinery than we are proposing to build. A mammal running a fast, wrong, useful model of a conspecific's internal state is an existence proof that the problem is tractable at some fidelity. The research question is which fidelity is enough for which purpose.

The furthest ground: exocortex, and a model that acts as you would

The long-horizon framing that motivates the programme is 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 — a digital clone in the strict sense of a behavioural and cognitive model, not a copy of a person.

This is where the field's vocabulary gets ahead of its evidence, and the project's position is that the vocabulary should be held loosely and the measurements held tightly. The distance from a 414-parcel dynamical model that cannot yet beat a persistence baseline to anything in this paragraph is not a matter of scaling. It is a sequence of unsolved problems, several of which we can now name precisely because we tried and failed at the first one.

What would change our mind

A vision section without falsifiers is marketing. Four things, any of which would substantially undercut the programme:

• Structured regional state does not beat a pooled vector at matched capacity. This is the pre-registered hypothesis of the current run. If the answer is no, the paper's first differentiator is wrong and the architecture is more complexity than the evidence supports.
• The anatomy prior does not resolve operator families. Already partially true — 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 for this programme 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.