Brain Neural Dynamics
The brain does not
do one thing
at a time.
Thirty-two electrodes. Four hundred and ninety-six relationships between them. Every one of them changing by the millisecond.
Brain Neural Dynamics · Phase 1 · 2026
Brain Neural Dynamics
Thirty-two electrodes. Four hundred and ninety-six relationships between them. Every one of them changing by the millisecond.
The standing assumption
Microstates, hidden Markov models, state-space clustering: at every instant the method picks one state and sets the rest to zero.
The field's dominant method, as reviewed in 2026
The crack in it
Measured geometrically, the timepoints assigned to one microstate sit roughly equally close to several. The distributions overlap. Hard labelling forces an ambiguous instant into a confident answer.
What BND does instead
Probabilities are never normalised across networks. Two networks can share Pz and both be fully on. There is no softmax, no partition, no winner.
And no fixed count: the number of networks emerges from the data
The second assumption we dropped
Every method picks a window length. On a one-second epoch, a 250 ms window has already erased the dynamics you came for.
Tested against the alternative
Worst-recovered true network, Pearson r · the exclusive model used 2× to 16× the free parameters
And you can watch it happen
Every network keeps its colour and its identity across the whole epoch. Scrub the millisecond, isolate a network, watch two of them share an electrode.
Sensor space only. The head is scenery, not a claim about where anything is.
What it cannot do yet
The count rests on a threshold calibrated against noise, and that threshold moves. One dataset, four honest calibrations:
test/cut.mjs · published, not buried. This is Phase 2's first job.
Brain Neural Dynamics
Every network here is a component recovered by a non-negative CP fit. Nothing on this stage was drawn by hand.
Every number on screen comes from a test that can be re-run on demand, against a hidden ground truth. The recovery and benchmark results hold. The count does not, and the film says so.
Not source localisation. Not a claim about where in the brain anything sits. BND infers relationships between scalp electrodes, resolved in time and frequency, and the anatomy on screen is reference only.
Three overlapping networks recovered at r = 0.995, 0.991 and 0.852, at the right sites and the right frequencies, with nothing found in matched noise. Counting is calibrated, not solved.