Engineering
How the models are built, what the experiments establish, and where the implementation still falls short.
SC-WBD connects regional dynamics to recordings through explicit geometry, observation models and clocks. These notes explain the implementation choices and the experiments that tested them. Start with a runnable release, then follow the topics relevant to your work.
Start with a released model
bWIN EEG 001 is the smallest self-contained starting point: a source wheel, trained weights and a synthetic example for 24-to-64-sample EEG forecasting. The release documents the exact channel order, preprocessing and inference recipe, alongside statistical and community-model comparisons. Its installed CPU example reproduces without the development repository or training data.
For the integrated model, use the SC-WBD-004 checkpoint and source repository. The results page summarizes each release's training and measured limits. bWIN and SC-WBD are separate architectures; check the input contract before comparing or reusing their outputs.
Representing state and observations
-
01
Orientation and spatial resolution
On the model's 400 cortical parcels, one scalar per parcel carries 32.1% of the EEG observable; three oriented components carry 83.4%. The experiment separates gains from orientation and spatial subdivision.
-
02
Combining observations across montages
How supports, units and component frames describe two instruments observing one state. Restriction and prolongation record what a transform loses and what it introduces.
-
03
Registering fMRI into parcel space
The EPI–T1w–template chain, evaluated on 55 runs from ten people. Coverage ranges from 0.89 to 1.00; per-run masks keep uncovered parcels out of the likelihood.
Fitting and evaluating predictions
-
04
A reproducible EEG forecasting comparison
bWIN, AR16, VAR4 and two community architecture adaptations on the same 25-person test, plus a separate five-person WBD comparison. Training histories, paired intervals and context controls show where the learned dynamics help and where simpler models perform better.
-
05
Training a conditional posterior
Three normalizing-flow failure modes, their signatures in the metrics, and the effect of separating posterior and model learning rates. Conditioning and calibration require their own measurements.
-
06
Point accuracy and uncertainty
Separating mean prediction from variance estimation to locate run 1's error. A Gaussian log score depends on both; report it alongside point error.
-
07
What the identifiability benchmark resolved
Five experimental designs in a three-region linear-Gaussian model. The registered fusion criterion never passes; other criteria depend on regime, and two cannot be evaluated with the chosen reference truth.
Implementation limits and reliability
-
08
Structured state versus a pooled representation
How the first implementation became the control arm of its intended ablation, and what that means for interpreting the measured result.
-
09
Where code and architecture differ
The documented gaps between the proposed architecture and executed code. Use this register to establish which mechanisms a result actually exercises.
-
10
Testing whether a check detects its failure
Examples of checks that stayed green while the quantity they were meant to test changed, and how to expose that blind spot.
-
11
Data paths, errors and provenance
Practical failures in file matching, exception handling and checkpoint provenance that affected what training and evaluation ran.
-
12
Defect and regression-check reference
A compact register linking known defects to the checks that detect their recurrence.
A practical review checklist
- Trace the dependency. A quantity described as conditional on an input must depend on that input in the executed expression.
- Exercise the failure. Apply the error a check is meant to catch and confirm that its output changes.
- Match the claim to the experiment. Record the cohort, preprocessing, baseline and uncertainty; a simulated example establishes only what its assumptions allow.
- Replay the artifact. Install the distributed package and reproduce an example outside the development checkout.