Foundation model
SC-WBD: an integrated whole-brain foundation model across modalities, scales and dynamics
One subject's brain as a 414-parcel dynamical system that EEG, MEG, fMRI and stimulation all supervise at once — each modality constraining only what it observes, with the rest marginalized rather than imputed.
SC-WBD models a single brain as a 414-parcel dynamical system and learns it from whatever was measured. EEG constrains the fast electrical state through a lead field built on that subject's head geometry; fMRI constrains a slow haemodynamic state in the same parcel space; TMS and tFUS enter as an E-field on the same grid. One state, one set of dynamics, four instruments looking at it.
The engineering problem is that these instruments disagree about time, space
and units. EEG samples at milliseconds and BOLD at seconds. EEG sees a current
dipole moment in Hz·m; BOLD sees percent signal change. A
subject scanned with one modality has no measurement of the others. SC-WBD
carries a state that is explicit about all three — per-region structure,
named clocks, typed ports — so a partial observation supervises the part
it observes and nothing else.
400 cortical + 14 subcortical
EEG · MEG · fMRI · TMS/tFUS
captured by a 3-vector moment, vs 5.6% scalar
each with a test that makes it fire
all five claim gates could_not_run
What exists today
Stated at the granularity someone could check. Every row below is a thing you could run.
| Component | State | What it is |
|---|---|---|
| Schema and compiler | built | Typed contracts for units, clocks, frames and transform lineage, with eleven refusals (R01–R11) that fail closed rather than warn.scwbd/schema/refusals.py A twelfth, R12, is local to this repository and fires at checkpoint emission. |
| Anatomy prior | built | 414 parcels — 400 Schaefer cortical, 14 Tian subcortical — carrying receptor density, intrinsic timescale, myelin and thickness. Partitioned into nine families by a pre-fixed rule under a spin null.1The rule was fixed before the test ran: ship the finest candidate partition in which every pair of families separates. Yeo-7 separated 6 of 21 pairs and was rejected. The surviving split is binary over cortex — cortex_unimodal (138 parcels) and cortex_association (262) — plus seven subcortical families of two parcels each. |
| Dynamics backends | built | Six generative backends — Wilson–Cowan, Jansen–Rit, reduced Wong–Wang, Stuart–Landau, Kuramoto, Linear–Gaussian — plus an equal-capacity learned control, interchangeable by one config key.2Twelve backends are registered in total once the four engineered per-family backends are counted. Only three of the six generative ones are typed mechanistic in the code; Stuart–Landau and Linear–Gaussian are effective and Kuramoto is functional. The distinction is in the source and we keep it. |
| Simulated corpus | built | 37,888 trajectories over 414 parcels at 125 Hz, generated from five of the backends.reports/run2_corpus.md |
| Real EEG corpus | built | 109 PhysioNet participants, 64 channels, split 71/11/27 and verified participant-disjoint. This is real human recording and is kept strictly separate from the simulated corpus.3The repository is emphatic about this and so are we: a simulated corpus can never be evidence that the model has learned anything about biology. Conflating the two corpora would be the most basic error available to this project. |
| TMS impulse-response path | partial | A validated E-field solver joined to the latent dynamics as an additive drive. The field computations pass their gates; the mapping from drive to neural response is explicitly flagged unvalidated in code, because no checkpoint has been trained on perturbational data.scwbd/intervene/impulse_response.py |
| Trained model that beats baselines | does not exist | Run 1's artifact lost. Run 2's pilot is training now. Nothing on this site should be read as a claim that we have one. |
| SC-WBD-002 | training | The first model on real anatomy — and, as it turns out, trained
on simulation alone.4Found at step 6166 of 8700, while the run was
still going. Run 2 renamed its training stages; the trainer gates
measured-data losses on the previous run's stage names, so
the gate never opened. The real-EEG loader was built and its split
fingerprinted, and no stage ever took a gradient on it — for nine hours the
only loss field emitted was sim_forecast_nll. Nothing raised,
because an unmatched stage name means "no real loss" rather than "unknown
stage". Two further mechanisms were inert for the same reason: per-stage
gradient permissions all fell back to the wildcard, and no individualizer was
ever built, so the stage named T1_individualisation ran 900 steps
of ordinary simulator training. 2.5M parameters over 414 parcels with
per-family engineered backends, heterogeneous per-family state widths, and
state-dependent predictive variance.5This entry previously said "a
ragged state layout (52% of the padded plane removed)". Both halves were
wrong. Run 2 ships the padded layout — the checkpoint records
layout: family_padded — with 47.34% of the
(414 × 59) plane being pad, because two hippocampal parcels of width
2 set the width for all 414 regions. The ragged layout is built and tested but
is not what these weights use, and the 52% figure was computed against an
11-family partition that no longer exists. Distinct from
001-beta in a way that is not a judgement call: the run-1
checkpoints refuse 414-region input, so the two do not share an
input space and cannot be compared. |
| Impulse response | measured | Two TMS coil poses produce measurably different predicted EEG, and a 200-permutation shuffled-normal null puts the real contrast above every permutation (p = 0.005): orientation carries the difference, not field magnitude. The criterion was committed while the checkpoint directory was still empty. |
Published artifacts
Everything below is on the Hugging Face Hub, public, with a card that states what it is and what it is not. The licences differ per artifact because the data underneath them differs — that is computed per artifact, not asserted once.
| Artifact | What it is | Licence |
|---|---|---|
| anatomy-prior-414 | The 414-parcel group-average prior: connectome, geometry, cortical normals, folding coherence. | Not NC-SA — it omits the regional maps, which are the one production asset carrying a non-commercial share-alike term. |
| sim-corpus-414-subset | Simulated trajectories over the same 414 parcels, for anyone reproducing the training curve without the measured corpus. | Generated by this repository. |
| scwbd-001-beta | Run-1 weights, published as a negative result: it loses to copying the last observed sample forward. Structurally the control arm of our own ablation. | Inherited from the corpora it was trained on. |
| scwbd-002-pilot | Run-2 weights — the treatment arm, family-indexed
heterogeneous regional state. Published as a negative result:
it loses to every baseline on both columns, NLL and MSE.6NLL
3.179 against 2.045 for the best baseline; MSE 36.27 against 4.53. Every
paired participant-clustered interval excludes zero. Run 1 at least won the
conditional mean — run 2 does not, so the "the loss is all in the variance
channel" diagnosis does not transfer.
And 3.179 flatters it. SC-WBD is scored on target/s with
the Jacobian folded into the log-variance; every baseline is scored on the raw
target. NLL_scaled = NLL_raw − log s, and mean(log s) =
0.5694 on this fold — roughly 17× the 0.035-nat spread across the three
non-trivial baselines. In their units the NLL is ≈ 3.75 and the gap ≈ 1.70
nats. No verdict changes; every interval moves further from SC-WBD.
Two more things the model card carries. subject_specific_ar is
bit-for-bit identical to ar16 — the participant-disjoint split
leaves no test participant with a fitted model, so the thesis's hardest
baseline never ran, while its own describe() reports
fallback_subjects=0. Read the table as five
comparators. And individualisation cannot be measured on this holdout at all:
no held-out person has a fitted person effect, so the between-participant
spread of the applied θ shift is exactly 0.000e+00. That is a
property of the split, not a defect of the run. The number that explains it: it scores
0.565 on simulation and 3.179 on measured
EEG, because five of six curriculum gates were wrong and no gradient was ever
taken on a recording.
And that is the smaller half. 88.7% of the model could not receive a
gradient at all — 2,231,447 of 2,516,530 trainable parameters,
including the entire family-indexed regional model that is the treatment arm's
whole thesis.7The regional modules were renamed local →
family_local, residual → family_residual,
readout → family_readout when the family-padded
architecture landed. The source cards still grant local.*,
residual.*, readout.*, and
fnmatch("family_local.ports.out_proj.weight", "local.*") is
False. An unmatched glob is not an error — it is an empty
permission set, which is a legal permission set. The 285,083 parameters that
were reachable are enough to fit something, so the loss fell, the run
finished, and five separate audits passed over it. Mechanism and measurement
agree independently: those exact modules are bit-identical across every
consecutive pair of the five stage checkpoints. So the negative result does
not say heterogeneous regional state fails to help. That part of the model was
a random initialisation participating in the forward pass for 8,700 steps. |
Inherited from the corpora; the anatomy prior's terms apply. |
Run 2 is done and listed above with its measured result, which is a loss. That was the commitment made here while it was still training — "it will be listed with its measured result, whatever that result turns out to be" — and the result turned out to be worse than run 1's.
Who this is for
Who it is built for, in order of how directly it is usable today.
- Researchers building multimodal brain models. The compiler, the refusal set and the identifiability machinery are usable now and are the part we would most like attacked.
- People designing non-invasive interventions — TMS, tFUS — who need a feasible-set discipline rather than an optimiser that will happily recommend something unsafe. The safety gating exists and refuses by default.8A plan declaring intent to drive real hardware or to be applied to a person is refused unless a record exists that a preliminary review occurred with an approving outcome. A valid authorisation record is necessary and explicitly not sufficient, and the gate does not open when a calendar date passes — there is a test that runs the clock forward to 2027 to prove it.
- Clinical translation. Not yet — this is a modelling and engineering programme, not a medical device.
The paper
The full technical description, built from LaTeX source in this repository rather than shipped as
a stale PDF.9Built with make paper, which runs
tectonic over
paper/sc_wbd_frontiers.tex. 29 pages. The source is the artifact;
the PDF is a build product.
This checkout has no git remote configured
(git remote -v is empty), so I could not link the repository or
verify a canonical URL. Every “view source” link on this site is currently
rendered as a plain path rather than a dead hyperlink.
Set SITE_REPO_URL in site/build.py (or export it
before make site) once the public repository exists, and the
provenance chips throughout the engineering essays become links.
Video
Two short pieces, rendered from code so that every figure in them comes from
the same source as the figures on this site.10Built with
Remotion under video/,
rendered locally with make video, and uploaded to object storage
with make video-upload. Media is never committed to the repository —
the compositions are the source, and the MP4s are build products.
Two videos are rendered and uploaded to R2, but the bucket is not served publicly, so there is no URL to embed here. Making a bucket world-readable is a hosting posture decision, so it was left to a human rather than taken on your account.
To publish them:
npx wrangler r2 bucket dev-url enable scwbd-media
then rebuild with SITE_MEDIA_URL=<the r2.dev URL> make site and this block is replaced by the players.
Alternatively attach a custom domain to the bucket, which is the better option for anything long-lived — r2.dev URLs are rate-limited and not intended for production traffic.