Cortex Laboratory — methods

Mathematical model and verification scope

Finite disk stimulation benchmark: an experimental, separate half-space configuration drives the same human membrane solver with the verified finite disk potential. Whole-cell coverage and the conservative 25 µm plane/contact guardrail are enforced. Recording physics, electric-field gradients and firing-threshold convergence are not established for this benchmark.

The laboratory uses the published human temporal active-neuron model for every experiment. The cell shares the implant world and calculates extracellular potentials from total transmembrane currents. Numerical checks establish agreement for stated mathematical cases. This implementation has no independent experimental validation.

One coordinate system

Positions use micrometers in a right-handed frame: pia is Z = 0; positive Z points into cortex. Rendering converts every structure uniformly to millimeters. Implant rotations apply X, then Y, then Z, followed by translation. Utah rotates about its tissue-facing substrate center; the single N1 segment and Neuropixels shank rotate about their distal end/tip. The scale bar measures distance in the camera-target plane; the coordinate inspector gives exact model coordinates and soma distances.

The Utah 1.5 mm reference comes directly from the existing device catalog. Its 100 physical shanks and 96 nominal channels are separate fields. The chosen sensing sites are hypothetical ideal probes, not an invented wiring map. Unknown exposed areas and channel assignments stay null. Neither substrate displacement nor electrical contrast is modeled.

Single thread and shank

Load one Neuralink N1 thread or one Neuropixels 1.0 shank. These are recording-only adapters, not full implants or surgical plans. Select two geometric sites to calculate ideal extracellular signals from the same human cell. Synaptic input or intracellular current clamp drives the neuron; neither is stimulation through these devices. Current implant edits invalidate prior results, while loading a curated example restores its complete Utah configuration.

N1 geometry follows Neuralink’s 2024 PRIME update and the DJ Seo engineering interview; it is distinct from later 128 × 8 descriptions. Neuropixels geometry follows the manufacturer’s 1.0 datasheet. Neuropixels 1.0 is a non-human research probe: placing its geometry beside a human cell here is a mathematical comparison, not a claim of clinical suitability.

Use Contact detail for the selected pair and Whole implant for the entire segment/shank. Geometry stays at physical scale; selection crosshairs and electrical diamonds are enlarged guides. Contact depths and distal tip depth are distinct. Sites outside the slab are labeled; the electrical model remains an infinite homogeneous conductor, not layered cortex/CSF. Finite electrode area, probe-induced displacement, insulation contrast and flexible-thread deformation are not solved. Exported results identify adapter revision, catalog revision, local and transformed coordinates, scope and sources.

Membrane and extracellular currents

For compartment i, C is total capacitance, Iᵢₒₙ includes leak and active ionic currents, Gᵢⱼ is axial conductance and Vₘ = Vᵢ − φₑ. Positive membrane current points outward. The solver uses mV, ms, nA, nF and µS, with explicit conversions from membrane area and resistivity.

Cᵢ dVₘ,ᵢ/dt + Iᵢₒₙ,ᵢ + gₛ,ᵢ(Vₘ,ᵢ − Eₛ)
  = Σⱼ Gᵢⱼ [(Vₘ,ⱼ + φₑ,ⱼ) − (Vₘ,ᵢ + φₑ,ᵢ)] + Iᵢₙⱼ,ᵢ

Iₘ,ᵢ = Cᵢ dVₘ,ᵢ/dt + Iᵢₒₙ,ᵢ + Iₛᵧₙ,ᵢ
Σᵢ Iₘ,ᵢ = Σᵢ Iᵢₙⱼ,ᵢ

The human cell has sealed terminals and a prescribed synaptic conductance step on a fixed physical apical region. Synaptic kinetics and placement are assumed inputs.

Human temporal active reference

The original reference and default target is specimen 571654895, a human middle temporal layer-2 spiny cell from Mosher, Wei et al. (2020). The reconstructed dendrites retain native SWC taper, section area and axial resistance. The published directed processor replaces the reconstructed axon with two 30 µm × 1 µm active sections. The source SWC contains a longer reconstructed axon, which is not the simulated axon. Arbitrary placement in this slab does not make the cell a motor-cortex specimen.

Nine mechanisms—NaV, Ih, Im, K_Pst, K_Tst, Kv3_1, Ca_HVA, Ca_LVA and SK—plus calcium dynamics and leak use the published fitted distributions. Human morphology and fitted densities do not imply measured human channel kinetics: sodium, potassium and M-current kinetics include mouse, rat and bullfrog sources. The temperature is fixed at the source model's 34 °C.

The membrane equation above includes all active ionic currents alongside capacitive and synaptic terms. Sodium uses the original twelve-state Markov model; other gates use their specified exponential state updates. Voltage uses channel states from the solve, followed by state updates. Exported currents and displayed end-state probabilities therefore have an explicit sampling-phase distinction. Shared zero-area branch junctions carry axial current but have no membrane, channel or recording source.

Initialization follows the source parameter JSON at −80 mV, then the declared no-drive conditioning period. The example simulation instead starts at −70 mV; that discrepancy is recorded. The port also preserves an upstream NaV initialization-equation discrepancy. Frame zero is the conditioned state, with residual drift reported by no-drive reference cases; exact equilibrium is not assumed.

Upward crossings of −25 mV are detected at internal solver steps and interpolated in time. Soma events and earliest local compartment crossings are distinct. Extracellular pulses can produce local dendritic crossings without a soma event; crossing location alone does not establish selective stimulation or a biological recruitment threshold.

The original model was fitted to surgical-slice whole-cell observations; paper Figure 3 reports training error. No held-out biological validation has been performed here. Source morphology, fitted parameters and reuse license are retained. The Allen Institute license permits noncommercial redistribution and includes a commercial redistribution restriction.

The extracellular point-source solution in an infinite isotropic conductor is φ = Σ Iₘ/(4πσr). The line-source option integrates 1/r uniformly along each cylindrical compartment and uses a point source for the soma. Tissue potentials are referred to infinity or differenced as A − B. The code rejects selected sources or sensing points within 25 µm of the modeled membrane instead of hiding the singularity with a distance clamp. That rule does not prove accuracy outside the excluded region.

Three distinct interventions

• Synaptic recording: a conductance event produces membrane currents and an extracellular signal. Total transmembrane current sums to zero.
• Intracellular current clamp: current enters the soma; an explicit bath return contributes both an applied field and a recording artifact. Net membrane current equals the injected current.
• Extracellular polarization: two ideal point contacts deliver equal and opposite biphasic currents. The resulting extracellular gradient drives the same cable. A uniform potential offset cannot alter its membrane response.

The neural contribution, applied-source artifact and their sum remain separate. Stimulation and recording cannot use the same ideal source point because its potential is undefined there. Charge per phase is descriptive. Exposed contact area is unknown, so geometric charge density is not supplied.

Reading the electrical view

Four explicit legends distinguish membrane voltage (mV), stimulating-contact outward current (µA), applied extracellular tissue potential (mV), and the selected ideal recording-contact component (µV). Neural potential comes from the forward sum of total membrane currents; applied potential comes from stimulating contacts or the bath return; total is their sum. No neural or total tissue volume is inferred from the two recording traces.

Stimulating spheres and recording diamonds are enlarged visual guides at the solver's exact contact coordinates. Their labels identify the ordered source/return or recording pair, physical contact ID, current or voltage, and polarity. Positive source current is outward into tissue (anodic); negative is a sink (cathodic). Source A leads cathodic, B supplies the equal/opposite local return, and the signs reverse in the second phase. Each recording marker remains an individual infinity-referenced sample when the plot is switched to A−B.

Applied slices use the same transformed point sources, conductivity and exported interval-average drive as the experiment. Choose XY, XZ, YZ or three planes, offset their center from the soma, adjust extent/opacity, and select fixed or automatic symmetric scales. Automatic tissue scale uses the largest visible sampled transfer times full pulse amplitude; recording auto-scale uses both contacts over the complete run. Thus neither scale changes during rest. Saturated vertex/contact counts are explicit. Neutral color means zero relative to infinity; blue and red indicate sign, not electric-field magnitude or direction.

Each plane has 25 × 25 calculated vertices with linear display interpolation. Both triangles of any cell intersecting a source-exclusion sphere are removed, even if no vertex lies inside it. Coarse holes can be larger than the stated sphere. No singularity is hidden by regularizing the distance. Tissue layers, pia and implant bodies do not alter this unbounded homogeneous solution. No electrode surface voltage or impedance is modeled. A new volumetric or isopotential view is deferred until it adds useful, independently checked information beyond these inspectable slices.

Membrane voltage is at the retained sample time; drive and artifact average only the preceding internal timestep. The replay cursor, readouts, traces and slice values all use that same sample. Phase buttons jump to actual retained negative/positive samples; previous/next stepping can inspect individual intervals. Missing pulse phases at coarse output sampling are reported. Slow replay cannot recover samples that were never retained.

Reproducible examples

Select a named example, read the reset disclosure, load it, inspect placement, then run. Each is a complete immutable configuration; loading clears current results and retains one pinned baseline. Pin the shifted array result and load its unshifted or polarity comparison to inspect the changed configuration fields, soma event times and recording metrics. Exact configuration and result exports are available in the laboratory.

The shifted-array candidate uses the published human cell, 40 µA, 1000 µs phases, 100 µs gap, translation [100, −50, 0] µm, stimulating r4-c4/r4-c5 and recording r3-c3/r3-c6. Its 0.005 ms / 40 µm calculation produces one soma event at approximately 5.381823 ms; the same pulse at [0, 0, 0] produces none. Reversing the ordered source pair produces a later soma event, approximately 6.657215 ms. These strong idealized inputs also produce very large hyperpolarization; none are safe-device recommendations.

All nine human-cell examples were rerun from exact configurations; measured reports and SHA-256 source/configuration fingerprints are retained in the example verification record. Original native NEURON/LFPykit fixtures cover the short intracellular and 200 µs local-crossing examples. Separate new native fixtures verify both sustained current-step examples, including all saved physical-membrane voltages/currents and ideal recordings; they pass the same numerical agreement tolerances. The numerical-reliability study adds a native comparison for the shifted 1000 µs candidate at 10 µm / 0.00125 ms. That comparison does not independently verify the default 40 µm shifted/reversed examples or every permitted placement. Original model fitting, numerical verification and independent biological validation are distinct.

Sustained current-clamp steps

The 200 ms and 500 ms patch-clamp examples inject +0.3 nA (300 pA) at the soma and allow repeated action potentials to emerge from the original channel mechanisms. They produce nine and eighteen soma crossings, respectively; the latter includes seventeen during the step and one immediately after offset. Longer interspike intervals show model-specific adaptation. Count divided by pulse duration is an average count rate, not a constant instantaneous rate.

Active inputs support up to 1000 ms steps within recordings up to 1200 ms. Saved output may be explicitly spaced up to 0.5 ms to stay under the unchanged 64 MiB budget; integration and spike detection still use the specified internal timestep. Coarser output can miss waveform peaks. Use a shorter recording and finer output to inspect those. These are ideal current-clamp calculations without pipette/access resistance or acquisition electronics, not voltage-clamp experiments.

Previous/next soma event and step-onset/offset buttons navigate retained frames. Exact interpolated event times remain in the event table and results export. Replay supports up to 100 biological ms/s. A longer replay cannot restore unsaved waveform detail.

Refinement and remaining numerical uncertainty

A bounded study compares 0.005, 0.0025 and 0.00125 ms timesteps at 40, 20 and 10 µm maximum compartment lengths. Comparisons hold the physical experiment and probe locations fixed, changing timestep or mesh separately. The study checks section topology, membrane area, arc length, capacitance and integrated channel conductance before interpreting numerical differences. It reports spike timing, local crossings, complete soma/selected-probe waveform errors, neural versus artifact errors, conservation, threshold brackets, runtime and numeric output memory.

Local-crossing thresholds remain spatially sensitive and nonmonotonic across these meshes; passing an independent-solver comparison at one discretization does not establish convergence. Recording-source chords change with refinement, but recording is one-way: chord errors affect recorded potentials and cannot cause a stimulation-threshold shift. The discretized cable, channels and center-sampled applied field jointly determine threshold sensitivity. All tested event counts/timing and applied-artifact comparisons passed, while sharp-spike waveforms failed predeclared voltage-error gates; the shifted candidate also failed one neural-recording error gate. All six extracellular spatial-threshold comparisons failed the 2% envelope criterion. Failed predeclared gates remain visible in the refinement record; no solver or physical model was changed to force a pass.

Numerical reliability and response interpretation

The response panel separates membrane polarization, local voltage crossings and soma threshold events. The legacy exported spikes array retains the −25 mV upward-crossing detector and 2 ms counting refractory interval. It is an event list, not an independent action-potential classifier. A saved positive-voltage peak followed by a return below the detection threshold is reported as descriptive AP-like waveform evidence. Coarse saved sampling or a short recording can miss that evidence; its absence does not disprove an action potential.

The earliest local crossing and delay to the first soma crossing describe temporal order. They do not establish the biological initiation site or propagation: an extracellular field can polarize several parts of the neuron directly. A causal propagation claim requires additional pathway and intervention checks.

The original six extracellular spatial-threshold comparisons remain failures of their 2% gate. A stable event count is weaker evidence than a converged threshold or waveform. Fresh native-simulator boundary comparisons and finer offline discretizations are recorded in the numerical reliability report, including threshold brackets, waveform errors and unresolved criteria. No tolerance was relaxed to make a case pass.

The two-section 60 µm synthetic axon does not represent the full axonal arbor or distant passing fibers. The source cell is a human middle temporal layer-2 specimen; moving it into a displayed motor or sensory layer does not validate it for M1 or S1. Consequently these checks do not establish cortical-population selectivity, a minimum resolvable neuron separation, or a real-device activation radius.

Parameters and evidence

The following table describes tissue/display assumptions. Section capacitances, resistivity and channel densities come from the pinned human model parameter file and are retained in result bundles alongside configuration and units. Default integration is 0.005 ms, output sampling is 0.05 ms, maximum segment length is 40 µm and initialization conditioning is 100 ms. These are numerical choices, not empirical confidence intervals.

The cortical slab is 6 × 6 × 4 mm by default. Layer fractions are 0, 0.06, 0.16, 0.46, 0.50, 0.75 and 1 times thickness. These boundaries, including layer IV, are illustrative. The displayed CSF and white matter do not create different conductivities. Curvature, vessels, meningeal mechanics, neuronal density and excitatory/inhibitory populations are absent.

Verification, exports and performance

Reproducible tests check SI unit conversions, line-source quadrature, dipole decay, analytic RC and two-compartment responses, sealed-cable polarization, current conservation, gauge invariance, rigid transforms, charge balance, singularity rejection, deterministic runs and exact CSV/JSON agreement.

Historical kernel checks, retained offline rather than selectable in the laboratory, include independent NEURON 9.0.1 references for a uniform passive cable and the complete declared 48-compartment arbor. The arbor comparison includes all three protocols and a rotated implant with a pulse onset between timesteps. Across fourteen sampled frames per case, voltage agrees within 9 × 10⁻¹² mV and total membrane current within 4 × 10⁻¹³ nA. NEURON builds its own sections and solves their equations; shared inputs contain geometry and parameters only. Its five membrane-free soma connectors reproduce the declared radial-resistance approximation, not a resolved spherical soma.

LFPykit 0.6.2 independently computes point/line-source coefficients and recordings from those NEURON currents. Eight cases compare neural, applied, total and bipolar signals, agreeing within 3 × 10⁻¹³ µV at sampled times. The tested geometries do not activate LFPykit's near-source radius regularization. These are fixed numerical comparisons, not biological error bars or guarantees for arbitrary configurations. Generators, package versions, exact inputs and hashed fixtures are retained in the repository.

Active references independently import the source SWC into NEURON 9.0.1 and compile the original NMODL mechanisms. Checks cover held-voltage channel trajectories, no-drive and intracellular responses, extracellular pulses, changed placement, membrane currents and LFPykit recordings. All physical compartments are compared over the complete exported sampling grid, with internal-step crossing times retained separately. Predeclared gates, exact scope, fixture quantization, refinement results and achieved errors are documented in the repository's ACTIVE_MODEL.md and validation report.

Run node --test tests/cortex-active-reference.mjs tests/cortex-current-step-reference.mjs and node scripts/benchmark-cortex-active.mjs in the repository. The benchmark report records hardware, timings, resolution and convergence. The on-page timer measures the local numerical run, not rendering FPS. Cross-browser speed, full peak memory and GPU performance have not been measured.

JSON bundles include configuration, engine and catalog revisions, geometry, source registry, frame-major voltages/currents, separated potentials and diagnostics. Active bundles also retain fitted parameters, channel outputs and spike detection conventions. CSV exports the chosen signal component and every compartment current. Frame zero follows declared no-drive conditioning. Active output sampling may be coarser than integration: each exported drive/artifact value averages only the preceding internal timestep. No analog filtering, anti-aliasing or ADC is implied. The laboratory accepts human active-cell configurations using version 2. Historical passive version-1 files are rejected with instructions to load a human-cell example; they are not silently converted. These are experiment schemas, not general SWC, SONATA, NeuroML or NWB importers.

Primary sources and scope

1. Mosher, Wei et al. (2020): Cellular Classes in the Human Brain Revealed In Vivo by Heartbeat-Related Modulation of the Extracellular Action Potential Waveform

   Figure 3; STAR Methods all-active single-neuron models; repository specimen 571654895. DOI: 10.1016/j.celrep.2020.02.027. Evidence: Human middle temporal gyrus, layer 2 spiny cell; surgical slice. Original fitting/training comparisons; this project has no independent held-out biological validation. Channel kinetics include nonhuman sources.

2. Allen Institute (2020): Pinned human active model source and reuse license

   Commit 750e7840cc1a9eb32492c924e79609a85a893b8a; assets/models, assets/morphologies, assets/modfiles, LICENSE.txt. Evidence: Human morphology/fitted densities; mixed-species kinetic mechanisms. Specimen identifier 571654895 is not a hosted Allen neuronal-model identifier. The synthetic 60 µm axon is the published model convention.

3. Lindén et al. (2014): LFPy: a tool for biophysical simulation of extracellular potentials generated by detailed model neurons

   Sections 2.1–2.2, equations 1–6; section 5.2.4. DOI: 10.3389/fninf.2013.00041. Evidence: general theory; examples are not human M1. An accepted equation does not validate our morphology or absolute amplitudes.

4. Anastassiou et al. (2010): The Effect of Spatially Inhomogeneous Extracellular Electric Fields on Neurons

   Printed pages 1926–1927. DOI: 10.1523/JNEUROSCI.3635-09.2010. Evidence: passive theory; rat CA1 morphology examples. Not human motor recruitment evidence.

5. Wagstyl et al. (2020): BigBrain 3D atlas of cortical layers: Cortical and laminar thickness gradients diverge in sensory and motor cortices

   Results: Comparison of total and layer thickness maps; Methods: Shrinkage estimate. DOI: 10.1371/journal.pbio.3000678. Evidence: human; single postmortem donor. Our slab and layer fractions are assumptions, not an extracted M1 profile.

6. Eyal et al. (2016): Unique membrane properties and enhanced signal processing in human neocortical neurons

   Abstract and figure 1/3 captions inspected; full tables not verified. DOI: 10.7554/eLife.16553. Evidence: human temporal L2/3. Not used to parameterize our synthetic cell.

7. Koessler et al. (2017): In-vivo measurements of human brain tissue conductivity using focal electrical current injection through intracerebral multicontact electrodes

   Abstract; online 2016, issue 2017. DOI: 10.1002/hbm.23431. Evidence: human epilepsy cohort. 50 kHz observations do not validate our low-frequency conductivity assumption.

8. LFPy developers: LFPy official API

   PointSourcePotential, LineSourcePotential, RecExtElectrode, StimIntElectrode; retrieved as 2.3.9rc2 docs. Evidence: not applicable. The standalone LFPykit forward models are checked in fixed numerical cases; finite-contact and layered-conductor behavior remains unverified here.

9. NEURON developers: NEURON 9.0.1 geometry and extracellular mechanisms

   Stylized geometry; extracellular mechanism; CVode.use_fast_imem; pinned reference fixtures in tests/fixtures. Evidence: synthetic passive circuit; no biological specimen. Fixed numerical cases verify the declared equivalent circuit, not a physical spherical soma or biological predictions.

10. LFPykit developers: LFPykit 0.6.2 electrostatic forward models

    RecExtElectrode pointsource/root_as_point; PointSourcePotential; installed version 0.6.2. Evidence: synthetic geometry and independent NEURON currents. Eight fixed ideal-contact cases stay outside radius regularization; no finite exposed area, electronics or experimental validation.

11. Blackrock Neurotech: Utah Array specifications

    Specification table; catalog BTSD-0001 / utah-10x10-1p5 r1. Evidence: device geometry. Taper and tip exposure are reconstructed; unknown area and channel map remain null.

12. Budday et al. (2017): Rheological characterization of human brain tissue

    Sections 2.1–2.4; tables 3–6; figures 6–10. DOI: 10.1016/j.actbio.2017.06.024. Evidence: human postmortem tissue. Future constitutive reference; no insertion or mechanical model implemented.

Distinct published cells and populations

The library contains 16 unique human cortical specimens: ten pyramidal/spiny excitatory cells and six inhibitory interneurons. Fourteen are from Allen’s human EAP collection (Mosher, Wei et al., 2020); two basket-cell fits are from Multi-modal characterization and simulation of human epileptic circuitry. Each biological specimen is available once per population, including the selected target and excluded placements. Changing a position never creates another biological identity. Dataset, source layer and class filters determine the available pool; the selected target remains cell 1.

The EAP source retains its published directed 60 × 1 µm axon stub. Basket-cell fits use the source’s two 30 µm axon sections with diameters taken from the source axon. Their current same-specimen Allen reconstructions are available, but the exact historical morphology revision used by the epilepsy publication is not; these entries are explicitly labeled adaptations. The original 34 °C channel equations, −80 versus −78 mV initialization, section parameters and dataset-specific Kv2like rate behavior are preserved. Mixed-species kinetic origins and incomplete axonal extent limit biological conclusions.

Finger Atlas specimen 508378649 has both a measured morphology and a matched published active fit and is runnable here. The atlas’s other nine measured Allen trees and twelve H01 neurite references lack matched electrical models in this library. Their morphology remains inspectable in the atlas; firing is not invented. Procedural DRG, spinal, thalamic and receptor models are excluded from the measured-cell library. H01 and the Allen sources are not measured finger M1/S1 circuits.

The separate Cortex Population Laboratory defaults to 16 distinct cells in a cloud, with radii resampled in log distance from the requested ladder 1, 2, 5, 12, 20, 28, 40, 60… µm over 1–960 µm. Every additional soma has a seeded uniform random direction from stimulating contact A; seed 20261018 reproduces the angles. A regular 3D grid and concentric shells remain available. The origin can be contact A, pair midpoint or target soma. Only placement is generated; source morphology and fitted conductances are not synthesized. These are illustrative placements, not measured packing or connected networks.

The unchanged 25 µm exclusion measures distance to the nearest membrane, not merely the soma center. The default has 16 placed, 8 supported and 8 excluded. Unsupported cells remain gold with explicit reasons, null JSON responses and blank CSV traces. They are not classified as inactive. Each cell’s own full morphology is used for placement checks and 3D rendering. Close arbors can overlap. A numerical failure in a supported cell stops the batch with its identity.

Independent active cables receive the same prescribed field without synaptic or ephaptic coupling. Every supported cell retains its own compartment-voltage replay on the shared time grid; neighbors use Float32 replay storage while the target retains full channel histories. Play, pause and scrub animate all supported neurons with the same −100 to +40 mV scale. Unsupported placements stay gold. The work budget remains 200 million node-steps including conditioning, with at most 64 placements and no reuse of the available 16 specimens. Replay retention is included in the capacity estimate and capped at 128 MiB of additional voltage arrays. The solver timestep and morphology resolution are unchanged. Neural recordings sum across supported cells; the shared stimulus artifact is added once.

The upper count is limited by available unique specimens, 200 million node-steps including separate conditioning of every cell, 16 MiB compact output and 64 MiB detailed output per cell. At 40 µm / 0.005 ms / 40 ms, all 16 fit the work budget; finer or longer runs reduce the count. Detailed histories are retained for the target plus at most one working neighbor; exported neighbor output contains soma traces and response summaries. These bounds exclude object/GPU overhead and are not a browser peak-memory guarantee. Assets are compressed and loaded only for the requested models/resolution. The earlier 64-copy timing report describes a retired repeated-specimen mode, not the current unique-cell library.

The default cloud completed in approximately 45 seconds in a development Node process while verification was running; this is not a browser speed guarantee or biological real time. Reproduce supported-grid and cloud measurements with node scripts/benchmark-cortex-population.mjs. Distinct specimens differ in morphology and membrane parameters as well as angle and distance, so these maps do not isolate a distance effect or establish a continuous activation boundary.

Native NEURON checks cover current clamp and weak extracellular forcing for each imported specimen using matched SWCs, original NMODL, initialization and waveform. They compare soma waveforms, saved full-membrane snapshots and soma-event timing. These are numerical reproduction checks for declared cases, not independent biological validation or validation of every permitted population position. Existing sustained-patch and original reference checks remain. The six historical extracellular spatial-threshold convergence failures remain unresolved. Local crossings, somatic events, overshoot/recovery evidence, propagation, applied fields and neural recordings remain distinct.

Next scientific gate

Complete browser review and broaden active parameter/resolution coverage. Compare against independently selected electrophysiology with explicit liquid-junction correction and held-out observations. Broader distinct sourced cells, realistic axonal extent, biological population validation, multilayer finite-element fields and mechanics require separate modeling and validation. The existing legacy tools remain separate until their models meet the shared geometry and current contracts.

Return to the laboratory →