CompteWMNeuron¶

Module: sc_neurocore.neurons.models.compte_wm Source: Compte, Brunel, Goldman-Rakic & Wang, Cerebral Cortex 10(9), 910–923 (2000) DOI: 10.1093/cercor/10.9.910 Scope: excitatory pyramidal cell and incoming AMPA/NMDA/GABAA kinetics

Scientific identity¶

CompteWMNeuron implements the source paper's pyramidal leaky integrate-and-fire cell together with the incoming channel-state equations. It is not the complete 2,560-cell spatial working-memory network.

The public input pathways are deliberately separate:

external_spike=True increments the external AMPA gate;
spike_in=True increments the recurrent NMDA precursor. The historical argument name is retained for compatibility;
inhibitory_spike=True increments the incoming GABAA gate;
current is direct somatic current in nA.

An output spike resets the membrane and starts the refractory interval. It does not increment GABA: the former self-inhibition rule was not present in the cited model.

Equations¶

The membrane equation is

$$ C_m\frac{dV}{dt} = -g_L(V-E_L) -g_{AMPA}s_{AMPA}(V-E_{exc}) -g_{NMDA}B(V)s_{NMDA}(V-E_{exc}) -g_{GABA}s_{GABA}(V-E_{inh}) +I. $$

The Jahr–Stevens magnesium-unblock factor is

$$ B(V)=\frac{1}{1+[Mg^{2+}]\exp(-0.062V)/3.57}. $$

Incoming channel kinetics are

$$ \frac{ds_{AMPA}}{dt}=-\frac{s_{AMPA}}{\tau_{AMPA}}, $$

$$ \frac{ds_{NMDA}}{dt}=-\frac{s_{NMDA}}{\tau_{NMDA}} +\alpha x_{NMDA}(1-s_{NMDA}), $$

$$ \frac{dx_{NMDA}}{dt}=-\frac{x_{NMDA}}{\tau_x}, \qquad \frac{ds_{GABA}}{dt}=-\frac{s_{GABA}}{\tau_{GABA}}. $$

Presynaptic events add one to the corresponding AMPA, NMDA-precursor, or GABAA state before the continuous flow. The five continuous variables advance together with explicit midpoint RK2 at 0.02 ms.

Threshold detection is sampled after the step:

$$ V_{candidate}\ge V_{threshold} \Rightarrow V\leftarrow V_{reset},\quad t_{ref}\leftarrow\tau_{ref}. $$

The paper used firing-time interpolation. This implementation does not claim equivalence to that within-step timing scheme.

Source control-set defaults¶

Conductances use µS, voltages mV, current nA, capacitance nF, and time ms.

Parameter	Default	Source meaning
g_l	0.025 µS	25 nS pyramidal leak
g_ampa	0.0031 µS	3.1 nS external pyramidal AMPA
g_nmda	0.000381 µS	0.381 nS recurrent pyramidal NMDA
g_gaba	0.001336 µS	1.336 nS interneuron→pyramidal GABAA
e_l, e_inh	−70 mV	leak and inhibitory reversal
e_exc	0 mV	AMPA/NMDA reversal
c_m	0.5 nF	pyramidal capacitance
mg	1 mM	extracellular magnesium
tau_ampa	2 ms	AMPA decay
tau_nmda	100 ms	NMDA open-fraction decay
tau_x	2 ms	NMDA rise-precursor decay
tau_gaba	10 ms	GABAA decay
alpha_nmda	0.5 ms⁻¹	NMDA saturation rate
v_threshold	−50 mV	sampled threshold
v_reset	−60 mV	pyramidal reset
tau_ref	2 ms	pyramidal absolute refractory time
dt	0.02 ms	source integration step

These conductances select the paper's control-set pathways for one pyramidal cell. Network weights and the connectivity footprint remain external.

Python API¶

Text Only

from sc_neurocore.neurons.models.compte_wm import CompteWMNeuron

cell = CompteWMNeuron()
event = cell.step(
    current=1.0,
    spike_in=True,          # recurrent NMDA event
    external_spike=True,   # external AMPA event
    inhibitory_spike=False,
)
state = cell.get_state()

get_state() returns v, s_ampa, s_nmda, x_nmda, s_gaba, and ref_remaining. reset() clears those dynamic values while preserving every configuration field. Invalid mutable state, configuration, current, or native input fails without partial state mutation.

The complete batch API is:

Text Only

result = cell.simulate(
    currents,
    recurrent_nmda_events,
    external_ampa_events,
    inhibitory_gabaa_events,
    backend="auto",
)

Each event array contains only zero or one. The result contains six complete state traces, the sampled output-event vector, and all six final-state values.

Executed backends¶

The maintained dispatcher exposes real Python, modular Rust/PyO3, Julia, Go, and Mojo implementations. The 200,000-step committed benchmark records 812 events in every lane. Python, Rust, Julia, and Go are binary64-identical for the complete receipt; Mojo remains within 3.6e-14, below the declared 2e-10 tolerance. These are local loaded-host regression measurements, not exclusive-core production-speed claims.

Native API documentation lives with each declaration:

Python docstrings on the class, step, batch, reset, and state surfaces;
Rustdoc on the modular engine, PyO3 boundary, and independent safety kernel;
GoDoc on exported state, validation, step, reset, and service functions;
Julia docstrings on the public module/type/step/batch/reset surfaces;
Mojo ABI and helper doc comments at the exported declarations;
RTL interface and fixed-point-contract comments at the module.

Source, schema, and silicon evidence¶

The independent 1,024-step primary-equation receipt pins all six state values and the four output events at indices 276, 505, 736, and 971. Its canonical row digest is bc071ae0c0057bc23a5e4c99ee5bbee53d306f07cd154e01bafff32335a6e192.

Paired TOML and JSON schemas implement a nine-edge lowering protocol and match the hand model. The Q16.16 RTL preserves the complete enrolled 1,024-step event vector, keeps voltage within 0.35 mV and each channel-state error within the declared bounds, synthesizes in Yosys, and passes a depth-4 CVC5 bounded safety job.

This is H1 evidence only. It does not establish binary64 formal equivalence, firing-time interpolation, timing, PPA, device behavior, or production deployment.

Retained SC network modification¶

The 2,560-cell ring, columnar connectivity footprint, Poisson drive, persistent bump, distractor resistance, and network statistics are retained as the separately named SC-COMPTE-WM-NETWORK project modification. They are not deleted, folded into this scalar cell, or promoted by this cell's evidence. The SC modification has its own five-runtime behavior, paired-schema, benchmark, and bounded ring-connectivity RTL/Yosys/formal evidence.