A working studio for brain dynamics.

This studio simulates large-scale brain activity on your machine by integrating the equations while you wait. You can set a brain state, give it a drug, and follow the dynamics from one ion channel up to the whole cortex.

Overview

math deterministic: Deterministic differential equations govern the whole-brain network, the spiking microcircuit, the single neuron, and the simulated scalp projections. These mathematical formulations contain no artificial intelligence, meaning identical initial values and noise seeds yield identical trajectories every time. You observe direct numerical integrations rather than stochastic estimations or trained approximations.
AI where the models are: Artificial intelligence enters only when interpreting data or organizing experiments. Cajal uses specialized tools to orchestrate runs and explain traces, while pre-trained models such as LaBraM read recordings and describe what is in them. Machine learning never computes the biophysical state equations, leaving the generation of physiological dynamics entirely to classical mathematics.

Sections of the studio

: The control bench and the 85-region brain: set a state, give it a drug, run it, and compare treated against untreated.
: A blind test of whether LaBraM, trained on real human EEG, can tell four simulated states apart from their scalp projections.
: One patch of membrane: set the ion gradients, read the resting potential, inject current and watch it spike.
: Four guided experiments, each locked until you commit a prediction.
: Every word the studio uses, defined, and the line between what is mathematics and what is AI.

Neural-mass dynamics: Wilson–Cowan populations over 85 regions

The engine follows how active each population of cells is, not every neuron. Each of 85 regions pairs an excitatory group with an inhibitory one, and the regions drive each other across conduction delays: a Wilson–Cowan neural-mass model, which is the right altitude for rhythms and seizures.

Figure 1. Above, a spiking circuit run cell by cell: every dot is one neuron firing once, orange for excitatory and cyan for inhibitory. Below, the single number the whole-brain model works with instead. Both come from one simulation, integrated when you opened this tab.

Three scales: membrane, microcircuit, whole brain

Three models, one question at three sizes. The single neuron computes voltage from ion gradients and channel gates; the microcircuit runs Izhikevich cells spike by spike; the whole-brain model couples 85 regions across the synthetic connectome.

Figure 2. One membrane spiking, one circuit's population rate, and the mean activity of 85 coupled regions. The same question asked at three sizes, each by a different model in this studio.

State space: six parameters, twelve conditions, fifteen treatments

Six knobs set a state, and twelve tuned conditions set all six at once. A treatment is a set of multipliers on the equations, chosen to match what the drug does at its receptor: the direction and rough size of an effect, not pharmacokinetics.

Figure 3. The awake brain, the same brain with its inhibition lowered until it discharges, and that seizure given a full dose of a benzodiazepine. Three runs, one changed parameter and one drug between them.

Runs: identifiers and determinism

Every run keeps a short id and the same settings give the same numbers every time. The local service holds your forty most recent runs for the session, so any of them can be reopened or handed to the assistant by id.

Figure 4. The awake state integrated twice from the same settings. The dashed line is the second run, drawn over the first: it lands on every sample of it, which is what having a run id is worth.

Whole-brain simulation: coupled Wilson–Cowan equations with conduction delays

We model large-scale brain dynamics by coupling 85 anatomical regions across a synthetic structural connectome with axonal conduction delays. Within each parcel, an excitatory population interacting with an inhibitory one governs local rhythmicity. Pressing run integrates these coupled Wilson–Cowan differential equations forward in time, and the cortical surface animates as the solver advances, translating regional population firing rates into an evolving spatial pattern.

Working the bench

Set the state. Pick one of the twelve tuned conditions, or describe one in words and confirm the card Cajal proposes. Either way the same equations run across all 85 regions.
Give it a drug. Add one of the fifteen compounds and press compare with untreated. The bench runs the same condition without the drug and tables what changed.
Go down a scale. Press run cell by cell to drop the same state into a spiking microcircuit and see every spike.

Investigations: parameter sweeps and transition sharpness

Every experiment on this tab proceeds through three beats: you frame an expectation, test it against differential equations, and inspect the mechanism that drove the outcome. The run button remains locked until you register a prediction. Committing to a specific outcome before you observe the data changes how you perceive the evidence, marking the boundary between an experiment and a demonstration. Once the integration finishes, your prediction sits directly beside what the equations produced.

A sweep evaluates a complete simulation across a series of parameter values while holding every other variable and random seed constant. By tracking the largest single jump against what an even response would have produced, we obtain a measure of sharpness. A straight line scores 1, whereas a score above 3 indicates the system tipped sharply across a narrow interval rather than sloping smoothly.

Single neuron: Nernst, Goldman–Hodgkin–Katz and Hodgkin–Huxley

A neuron is a battery: pumps hold potassium inside the cell and sodium and chloride outside, and each gradient pulls the membrane toward its own Nernst potential. The Goldman–Hodgkin–Katz equation weighs those pulls by how permeable the membrane is to each ion and gives the resting potential. A Hodgkin–Huxley membrane then shows how voltage-gated channels turn that resting potential into an action potential. Change a concentration, a permeability, the temperature or a channel drug and both models update together. math Everything on the tab is integrated from the equations; no learned model is involved.

EEG benchmark: closed-loop discrimination with LaBraM

Can a network trained strictly on clinical human EEG recognise a brain state generated purely from differential equations? We synthesise four distinct neurological states with the whole-brain model, project the activity onto virtual scalp electrodes, and feed the output to LaBraM, an EEG foundation model trained on roughly 2,500 hours of real human recordings. Because LaBraM has never encountered this studio's synthetic engine, it cannot fall back on a memorised artifact — whatever it separates, it separates on the structure of the signal itself. This is a closed-loop falsification test: if the simulated biology is mathematically hollow, the evaluation fails, and the page says so.

The four stages

Define. An external language model designs four distinct neurological conditions out of the studio's tuned biophysical presets. Adjustments are bounded to half a step from baseline, which prevents the over-saturated dynamics where every state degrades into uninformative slow waves. The selected states and their expected electrophysiological signatures are locked into the table on the left before anything is executed, eliminating retrospective bias.
Simulate and project. The neural engine integrates each configuration across the 85-region whole-brain connectome for eight simulated seconds. A distance-weighted forward kernel then projects regional cortical currents onto 19 standard 10-20 scalp channels, incorporating a calibrated microvolt gain, pink 1/f background noise and a 1 µV trace of line-frequency hum; only cortical regions contribute. Set beside a boundary-element structural head model this projection is abstract, but it faithfully preserves the three properties the test turns on — spatial synchrony, broadband power and amplitude dynamics.
Decompose and encode. Raw traces are bandpass-filtered between 0.1 and 75 Hz, notch-filtered for line noise, downsampled to 200 Hz and partitioned into four-second epochs. We deliberately avoid average referencing: subtracting the spatial mean artificially erases the global phase synchrony that is characteristic of generalised seizures and slow-wave sleep — the very thing those states are made of. Each epoch is then processed through two parallel paths. LaBraM compresses the multi-channel time series into latent embeddings, mapping the signal's geometry using representations learned from human clinical records, while a deterministic classical pipeline extracts canonical band power, cross-channel synchrony, a burst index and spectral flatness.
Discriminate and score. Two independent raters evaluate the data blind. A fresh instance of the external language model reviews the tabular spectral metrics with every identifying line stripped out, names the underlying physiological condition and reports its own confidence score. Separately — and reported by nothing — we measure the geometric distance between each epoch's embedding and its true condition's centroid. The run is marked successful only if three things hold: LaBraM read every recording, its latent clusters separate the conditions better than chance, and the blind judge correctly identifies at least 3 of the 4 conditions.

Scope and limits

Every trajectory on this page reflects a synthetic connectome and an idealized mathematical model rather than a clinical measurement from an individual patient. Tipping points and seizure thresholds are intrinsic properties of these equations, so the studio describes simulated network phenomena without offering medical diagnoses.

Every word the studio uses is defined on the tab, with the line between what is mathematics and what is AI, and the technical specification.

Control bench

A treatment acts on top of a condition, so choose the condition first.

condition

treatment

dose · 100% is a typical clinical effect100%

knobs

excitability1.00

inhibition1.00

coupling1.00

noise1.00

adaptation1.00

arousal1.00

duration ms3000

excitability: Sensitivity of the excitatory population, the analogue of pyramidal-cell gain. Up makes the brain hyper-responsive; pushed far enough it discharges on its own.
inhibition: Strength of the local inhibitory population, the analogue of GABAergic interneurons. It is the network's brake; dropping it removes containment.
coupling: Gain along the wiring between regions: how easily activity in one region jumps to its neighbours.
adaptation: Activity-dependent fatigue, the analogue of slow potassium currents. It builds while a region fires and forces it to pause: the rise and fall of slow-wave sleep.
arousal: Brainstem drive, the analogue of noradrenaline and acetylcholine. High arousal breaks slow waves into fast waking rhythms; near zero, activity collapses.
noise: Random background input standing in for subcortical fluctuations.

1 is the awake baseline for every knob; 0 is off. A knob you touch overrides the condition's value; small steps (0.2–0.5) change a state, large ones saturate the model.

Treatment effect math

measure	untreated	treated	change

A treatment here is a set of multipliers on the model's parameters, chosen to match what the drug does at its receptor. It shows the direction and rough size of an effect, not a dose-response or a prediction for any patient.

Brain view math

nothing loaded

nothing loaded yet

resting tissue01excitation

—

how to read the brain

What this is. A cortical surface rendering: a model cortex with each region shaded by its simulated firing rate, played forward through the run. Surface renderings like this are a standard way of presenting results in neuroimaging and computational neuroscience, used to study the spatial and temporal pattern of activity — which regions are active, in what order, and whether activity stays focal or spreads.

organ: The anatomical surface at full resolution. Hover anywhere for the atlas label of the parcel under the cursor.
depth: The cortex made semi-transparent, revealing the deep nuclei and the long-range wiring at their true depth; nearer structures are drawn larger.
slice: Orthogonal cross-sections: sagittal, coronal and axial through one point. Hover to name a region, click a pane to move the cutting planes. Nothing is highlighted until you hover it.
colour: Grey is quiescent tissue at rest, so a resting cortex still reads as anatomy. Excitation paints over it on the Turbo scale — blue through green and yellow to red, where a region saturates — and tracks the region's population firing rate directly.
playback: Quarter speed by default, so a travelling wave is slow enough to follow across the surface.
zoom: The − and + buttons, or pinch (ctrl-scroll) over a view, zoom a little; double-click to reset.
phases: Runs with stages are annotated: the bar under the scrubber marks each phase, the line under the title says what is happening now, and a white ring marks a region only while it is stimulated.
what it is not: A run shows the kind of dynamics a cortex has under a condition, on a synthetic connectome. It is never a measurement of anyone's brain; the full statement is on .

Readout · activity over time math

band power

–

delta

–

theta

–

alpha

–

beta

–

gamma

how to read the numbers

What this is. Two standard views of one signal: a time series and its band power. The line is the simulated activity of all 85 regions averaged into a single value at each moment; the bars beside it show how that signal’s power divides among the conventional EEG frequency bands. This pair is the routine readout in electrophysiology — waking, sleep, anaesthesia and seizure are distinguished far more by the shape and frequency content of such a trace than by any single instant of it.

mean activity: The spatial average of excitation across all 85 regions, moment by moment: the orange line above, named on the chart itself. In the modelling literature this quantity is the mean field. Low and steady means sparse, local firing.
amplitude: How big the swings of that average are, as a root-mean-square. Large amplitude means the whole cortex is rising and falling together rather than idling.
synchrony (cyan): How much of the regions' variation is shared: the variance of the mean field divided by the average variance of the individual regions. 0 is independent, 1 is lockstep. Wakefulness sits at 0.1–0.3, deep sleep and seizure above 0.6. This is a shared-variance ratio, not a Kuramoto phase order parameter.
dominant rhythm: The strongest single frequency, with the band holding the largest share of power. When activity collapses there is no rhythm to name, and the readout says so instead of naming one.
band power: The signal split by rhythm: delta 1–4 Hz (deep sleep), theta 4–8 (drowsiness), alpha 8–13 (relaxed wakefulness), beta 13–30 (alert processing), gamma 30–80 (local computation).
left / right: Mean activity per hemisphere, and the correlation between each region and its mirror image on the other side.
most active: The four regions with the highest mean excitation over the run, on the same 0–1 scale as the brain view.
first responders: Only after a stimulus: the regions that rose 3 standard deviations above their own baseline earliest, and how long after stimulus onset they did. This is a response latency.
entrained: Only after a rhythmic stimulus: regions whose power at the driving frequency grew, and by what factor, measured during the drive against before it. Stimulated regions are excluded, so entrainment here means a region that followed the rhythm rather than one that was given it.
notes: First readings generated by fixed rules, not by a model. The numbers above them are the evidence.

Membrane settings math

condition

01 · transmembrane gradients

K⁺ inside

K⁺ outside

Na⁺ inside

Na⁺ outside

Cl⁻ inside

Cl⁻ outside

02 · relative resting permeability (P / P_K, log scale)

P_K

P_Na

P_Cl

temperature

channel drug

03 · environment and stimulus

current step

step length

things to try

hyperkalaemia: Raise potassium outside. The resting potential depolarises, firing speeds up, then stops in depolarisation block once the sodium channels can no longer reset. The curve from investigation two.
GABA-A: Raise P_Cl. E_Cl sits near rest so the voltage barely moves, but the extra conductance shunts the injected current out before it reaches threshold.
tetrodotoxin: Block the sodium channels. Spikes vanish and only the passive charging curve is left.
hypothermia: Drop to 10–20 °C. Q10 slows the gates: broad, sluggish spikes and long refractory periods, with the equilibrium potentials unchanged.

Resting potential · Goldman–Hodgkin–Katz math

One patch of membrane, set to the controls. Dots show concentration, the width of each pore its permeability, and the weight of each arrow the resting flux — permeability × driving force (V_m − E_ion) — all relative to the largest of the three. Sodium is pulled hard inward but its resting channels are nearly shut, which is why potassium sets the resting potential.

ion	inside [i]	outside [o]	Nernst potential E	relative permeability P

The white marker is the resting potential. It sits just above E_K because the resting membrane is twenty-five times more permeable to potassium than to sodium. Raise P_Na, as the upstroke of a spike does, and V_m swings toward E_Na; raise potassium outside and E_K, and the resting potential with it, move toward threshold.

how to read the resting potential

What this is. A schematic patch of cell membrane above a voltage axis. The schematic shows the three permeant ions — potassium, sodium and chloride — at their intracellular and extracellular concentrations, with the channels that let them cross; the axis places each ion's equilibrium potential and the resting membrane potential the three of them produce together. This is the standard account of how a neuron holds a voltage across its membrane, and it is what every condition and channel drug on this tab acts on.

equilibrium potential (E): The voltage at which an ion's concentration gradient and the electrical force on it cancel, so there is no net flow. One value per ion, given by the Nernst equation and printed in the table.
membrane potential (V_m): The actual voltage across the membrane, the thick white mark on the axis. The Goldman–Hodgkin–Katz equation places it as a permeability-weighted average of the three equilibrium potentials, so the ion with the most open channels pulls hardest.
permeability (P): How readily the membrane passes each ion, stated relative to the others rather than in absolute units. At rest potassium's is much the largest, which is why V_m sits close to E_K.
driving force: V_m − E for one ion: how far the membrane sits from the voltage that ion alone would impose, and therefore which way that ion is pushed. The arrows in the schematic are weighted by permeability × driving force, so a large gradient through shut channels draws a small arrow.
what it is not: A textbook membrane with three permeant ions, not a particular cell type, and not a recording from any cell. The ion pumps that maintain these gradients are not drawn: a condition such as ouabain acts here by setting the concentrations it would leave behind.

Action potential · Hodgkin–Huxley math

the gating sequence

Trigger. The current step (the orange bar) charges the membrane until the potential crosses threshold, the value printed in the summary above.
Upstroke. Depolarisation opens the fast m gates; sodium floods in and pulls V_m toward E_Na, which opens more gates.
Downstroke. Near the peak the slower h gates shut sodium off while the n gates open potassium, driving the potential back toward E_K.
Reset. The h gates only reopen once the voltage is negative again. That is the refractory period, and it is what never happens in depolarisation block.

Reversal potentials are the Nernst values in the table above; the leak reverses at the Goldman potential. Gating kinetics are Hodgkin and Huxley's own, scaled by Q10 = 3 so that 37 °C runs at reference speed: a textbook membrane, not a particular cell type.

how to read the action potential

What this is. A membrane potential trace — the cell's voltage plotted against time while a current is injected — with the gating variables of the Hodgkin–Huxley model plotted underneath it. This is the conventional way an action potential is presented, and the Hodgkin–Huxley model is where the presentation comes from: the 1952 description of the squid giant axon that conductance-based neuron models still build on.

the white trace: Membrane potential in millivolts against time. Each sharp excursion is one action potential, or spike.
the orange bar: When the injected current is on, and at what strength. A step too small to bring the membrane to threshold produces no spike at all: the trace only charges and settles.
the dashed lines: E_Na, E_K and V_m (GHK) carried over from the panel above, on this voltage axis. A spike rises toward E_Na and falls back past rest toward E_K without reaching either, because both channel populations are open at once.
m, h and n: The gating variables, each between 0 and 1: the fraction of that gate population currently open. m opens the sodium channels, h inactivates them, n opens the potassium channels. Their order is the mechanism of the spike, stepped through above.
firing: The spike count and the rate in hertz during the current step, printed in the summary. A neuron that fires faster as the current rises is behaving normally; one that stops firing as the current rises is in depolarisation block.

Run the test

closed loop · the studio grades itself

1 · Definenot started
2 · Simulate and projectnot started
3 · Decompose and encodenot started
4 · Discriminate and scorenot started

The recordings math

19-channel scalp EEG all 19 channels one scale for all states

One strip per state: the trace LaBraM read, with its microvolt scale, the band powers beside it, and both readings — LaBraM's nearest-centroid pick with its geometric confidence, and the blind judge's pick with its own self-reported certainty. Each strip is drawn to its own scale so the rhythm stays legible; tick one scale for all states to compare amplitude. See the simulation opens the run a strip came from.

Nothing has run yet. Each state's recording appears here as the test computes it.

how to read a recording

What this is. A scalp EEG: the voltage at each electrode plotted against time, one row per channel, stacked front to back across the head. Every strip here is simulated — a whole-brain run projected onto 19 electrode positions of the international 10–20 system — but it is drawn and analysed the way a clinical recording is, because that is what the test turns on: whether a model trained on real human EEG can separate states generated by equations.

the rows: One per channel, named at the left in 10–20 notation. Six evenly spaced channels by default; tick all 19 channels for the full montage.
± µV: The vertical scale, printed on every strip in microvolts. Each recording is drawn to its own scale so its rhythm stays legible; tick one scale for all states to compare amplitude between states instead.
the gridlines: One second apart, labelled every other second along the bottom.
δ θ α β γ: The band power of that same recording: the share of its power in each conventional band — delta 1–4 Hz, theta 4–8, alpha 8–13, beta 13–30, gamma 30–80.
the two readings: LaBraM's nearest-centroid pick with its confidence, and the blind judge's pick with its own stated certainty. Each strip reports what was said about it; whether it was right is the benchmark result above.
what it is not: A simulated recording, not a patient's. The forward projection is a distance-weighted kernel rather than a boundary-element head model, so electrode positions are approximate and only cortex contributes.

Glossary

the words this studio uses

brain state: A stable regime the whole network settles into: alert, deep sleep, seizure. In the model it is one setting of the six knobs.
run: One simulation or one analysis, kept under a short id: wb- whole-brain, mc- cell by cell, nr- single neuron, eeg- a scalp recording, bench- a benchmark. Any run can be redrawn from History, and Cajal can look one up by id.
preset: A saved set of knob values that reproduces a named state. There are twelve.
knob: One of six numbers in the regional equations: excitability, inhibition, coupling, adaptation, arousal, noise. 1 is the awake baseline; 0 is off.
treatment: A drug or lab compound, written as multipliers on the model’s parameters and applied on top of a condition. It shows which way the compound pushes the network and roughly how far. The multipliers were written by hand from what the drug does at its receptor; this is not a pharmacokinetic model.
dose: How hard a treatment pushes, 0–150%. 100% is a typical clinical effect — or, for the lab compounds, a typical experimental one.
region: One of 85 nodes: 68 cortical parcels from the Desikan–Killiany atlas, plus 17 deeper ones — the subcortical nuclei, the cerebellum and the brainstem.
connectome: The wiring between regions, with a conduction delay on every link. Synthetic: a distance rule plus a handful of long-range pathways added by hand. Not measured tractography.
excitation: The firing rate of a region’s excitatory population, 0 quiet to 1 saturated. On the brain it is painted over the tissue grey on the Turbo scale: blue is weak, then green and yellow, red where a region saturates. Resting cortex keeps its anatomy, so colour always means activation.
mean field: The average excitation across all 85 regions, moment by moment.
synchrony: How much of the regions’ variation is shared: the variance of the mean field over the average variance of the regions. 0 is independent, 1 is lockstep. A shared-variance ratio, not a phase measure.
rhythm and bands: Delta 1–4 Hz (deep sleep), theta 4–8 (drowsiness), alpha 8–13 (relaxed wakefulness), beta 13–30 (alertness), gamma 30–80 (fast local processing). Band power is the share of the signal sitting in each.
sweep: Running the model many times over a range of ONE parameter, everything else held still, so you see the shape of the response instead of a single point. The Investigate tab is built on sweeps.
sharpness: How abrupt a response is: the biggest single step in a sweep divided by the step an even, straight-line response would have taken. 1 is a straight line; above 3 is a tipping point.
monotone: A response that only ever moves one way. When a sweep is not monotone, the same change helps and then hurts — usually the interesting part.
whole-brain model: A Wilson–Cowan neural-mass model: excitation–inhibition loops in all 85 regions at once, stepped forward in time.
microcircuit: A sparse random network of Izhikevich spiking neurons, for looking at a state cell by cell.
membrane potential: The voltage across a neuron’s membrane, inside relative to outside. About −70 mV at rest.
Nernst potential: The voltage at which one ion’s concentration gradient is exactly balanced by the electrical pull on it, so it stops flowing: E = (RT/zF)·ln([out]/[in]).
permeability: How easily an ion crosses the membrane, set by how many of its channels are open. At rest the membrane is most permeable to potassium.
Goldman equation (GHK): Where the resting voltage settles when several ions cross at once, each gradient weighted by its permeability. It explains why rest sits near E_K, and why opening sodium channels swings the voltage toward E_Na.
Hodgkin–Huxley model: The equations for a spiking membrane: sodium channels with two gates (m opens fast, h shuts slowly), potassium channels with one (n), and a leak. Their interplay turns a depolarisation into an action potential.
depolarisation block: A neuron held so far above its resting voltage that its sodium channels stay inactivated. Maximally excited, and completely unable to fire.
EEG: Electroencephalogram: electrical potentials recorded at the scalp, nineteen of them here. The studio reads real EEG files, and can make a synthetic one from a simulation.
forward model: The step that turns a simulation into a scalp recording: each of the 19 electrodes gets a distance-weighted sum of the cortical regions under it, plus pink noise and a trace of mains hum. A distance kernel, not a head model — rhythm and amplitude survive it, precise localisation does not.
epoch (window): Four seconds of recording, all 19 channels together, taken as one unit. The benchmark cuts every recording into epochs and scores them one at a time. The two words mean the same thing here.
LaBraM: Large Brain Model, a network trained on about 2,500 hours of real human EEG. Give it one epoch and it returns a few hundred numbers describing what that stretch looked like. It names nothing and diagnoses nothing.
latent embedding: Those few hundred numbers, read as coordinates: two numbers place a dot on a map, several hundred place a signal in latent space. Recordings that look alike land close together, which is what lets the benchmark ask whether simulated states can be told apart without ever being told what they are.
Cajal (agent): The AI assistant behind the chat, run by the Melchizedek runtime on a language model. It calls the studio’s tools and may report only what they returned.
blind judge: A separate instance of the language model. It reads only the automated analysis of a recording, never the state that produced it, and names the state.
confidence: Two different numbers, and the benchmark keeps them apart. The judge reports its own, 0–100%, and it is a self-assessment. LaBraM’s is measured instead: hold one epoch out, find which state’s centroid it lands nearest, and score how decisively it belongs there. Nothing reports that second number — it is calculated.

What is AI here, and what is not

the line, kept in one place

The studio keeps a hard line between the equations that generate the dynamics and the learned models that read them. Nothing on the left was trained on anything; nothing on the right computes a trajectory. That separation is what makes a result here possible to disagree with.

math: The simulations. The whole-brain model, the cell-by-cell model and the single neuron are systems of equations solved step by step on this machine. Same settings, same numbers, every time.
math: The treatments. Each compound is a set of multipliers on those equations’ parameters, written by hand from what the drug does at its receptor.
math: The analysis. Mean field, synchrony, band power, latencies and the EEG forward projection are classical signal processing. No model is involved in any of them.
AI: Cajal. A language model that decides which tools to call and explains what they returned. It may not invent a number.
AI: LaBraM. A network trained on real human EEG. It reads four seconds of all nineteen electrodes at once and returns a few hundred numbers describing that stretch. It names nothing: the benchmark tests only whether states that differ produce numbers that differ.
AI: The benchmark’s definer and judge. Language models that design the four test states and, blind, name them back from an analysis.
AI: The other pre-trained models. SwiFT and BrainGFM (fMRI), Geneformer and scGPT (single-cell), Cellpose-SAM (microscopy), each run when you hand Cajal a file. Any that is not configured falls back to classical analysis, and the result says which one ran.

Technical specification

disclaimer: cajal is a simulation studio and a testing harness for algorithms. It is not a clinical or diagnostic instrument, not certified as software in a medical device, and nothing in it should be read as advice about a person.
macroscale: Wilson–Cowan neural-mass formulation over 85 regions: 68 Desikan–Killiany cortical parcels plus 17 deeper nodes — subcortical nuclei, cerebellum and brainstem — on a synthetic distance-rule connectome with conduction delays. The wiring is generated, not measured tractography.
the solver: Plain TypeScript with no dependencies, integrated on the CPU. One copy of it serves both the agent’s tools and this page, so a run makes the same numbers wherever it was started.
geometry: FreeSurfer fsaverage cortex; cerebellum and brainstem from the MNI152 template; sections from the MNI152 T1.
microscale: Sparse random networks of Izhikevich spiking neurons.
single neuron: Nernst and Goldman–Hodgkin–Katz equations; Hodgkin–Huxley membrane with Q10 temperature scaling.
AI stack: EEG: LaBraM (active). fMRI: SwiFT, BrainGFM. Single-cell: Geneformer, scGPT. Imaging: Cellpose-SAM. All of them run locally on PyTorch. Agent: Cajal on the Melchizedek runtime (Gemini, or a local Qwen through Ollama).
architecture: Four local services: :8945 Python inference (npm run models); :8935 / :8975 the simulation tools and this page's event-stream relay (npm run serve); :4010 the Melchizedek agent service (npm run a2a); and the page itself (npm run dev). All four bind to loopback only.
where your data goes: Every simulation runs on this machine, as does LaBraM and every other pre-trained model: no recording, image or dataset you open is uploaded. The conversational agent is the exception. By default Cajal is served by Gemini, so the text of your questions and the tool results it reads are sent to that API. Point it at the local Qwen build (cajal_local) and nothing leaves the machine at all.