Wong-Wang source, runtime, and co-simulation fidelity

This page records the source boundary, independent equation trace, five-runtime parity, failure atomicity, fixed-point envelope, and controlled benchmark used to promote WongWangUnit.

Primary-source boundary

Wong and Wang (2006) reduce a recurrent spiking decision circuit to two NMDA gating variables. The Appendix reduction uses the transfer

[ \phi(I)=\frac{aI-b}{1-\exp[-d(aI-b)]}, ]

the coupled currents

[ I_1=J_NS_1-J_{cross}S_2+I_0+I_{stim,1}+I_{noise,1}, ]

with the symmetric expression for population two, and

[ \dot S_i=-S_i/\tau_s+(1-S_i)\gamma\phi(I_i). ]

SC-NeuroCore also carries the paper's AMPA Ornstein-Uhlenbeck current state, using an explicit external standard-normal sample for each population and physical step. The maintained scope excludes recurrent AMPA.

Primary source: Wong and Wang 2006. Author-lab implementation: pinned wong06.m.

Timestep and rate-order decision

The paper Methods state dt=0.1 ms; the pinned author script uses 0.5 ms. The maintained default is the paper value, 0.0001 s. This discrepancy is recorded rather than silently selecting the code value.

The author script stores an initial nu value before entering its indexed update loop. SC-NeuroCore instead returns the algebraic pre-update transfer of the complete current state on every call. This removes a one-index storage lag without changing the published recurrence.

Independent reference trace

src/sc_neurocore/neurons/reference_trace_data/wong_wang_appendix_euler_ou_doi.json pins a 256-step trace with:

• varied, deterministic stimuli to both populations;
• non-zero explicit noise samples;
• all four post-update physical states;
• both pre-update rates;
• the DOI, author-lab URL, and source commit;
• every first/final/minimum/maximum/mean feature.

tests/test_reference_wong_wang.py independently re-derives the equations without importing production equation helpers, checks every recorded feature, and then compares the complete production trace byte for byte. The canonical interleaved little-endian float64 digest is d39f219d3cd21d505c71749a1d9547d4cef550299f8e829bb2aa2a30d66daf44.

Numerical and atomic contract

Each scalar or batch step:

1. validates parameters, current state, stimuli, and the two samples;
2. evaluates both currents and both rates from the pre-update state;
3. evaluates two Euler gating candidates and two OU-current candidates;
4. validates all four candidates;
5. commits all four together.

No runtime clips an invalid gating candidate. Public Python batch state and C ABI output buffers remain unchanged when a complete result cannot be produced. Empty batches preserve all four physical states.

Executable parity matrix

Runtime	Executed surface	Enrolled result
Python	scalar reference and atomic batch	reference
Rust engine	modular PyO3 batch	six traces/four final states within 1e-12
Rust safety	standalone rustc --test module	scalar equations, validation, symmetry, reset
Julia	JuliaCall batch	six traces/four final states within 1e-12
Go	generated C-shared ABI	six traces/four final states within 1e-12
Mojo	exported shared-library C ABI	six traces/four final states within 1e-9

The configured test begins from s1=0.24, s2=0.11, noise1=0.01 nA, and noise2=-0.02 nA; changes every public parameter; supplies varied currents and samples for 128 steps; and compares all six traces plus all four final states.

Paired schema and Q32.32 co-simulation

The TOML and JSON schemas are structurally identical. Because the universal schema port accepts one scalar input per edge, one physical Wong-Wang update is serialised over six rising edges:

1. latch stim1;
2. latch stim2;
3. latch xi1;
4. latch xi2 and evaluate both transfer arguments;
5. evaluate both rates through the generated exprel path;
6. commit both gating and both OU-current states.

Gaussian generation remains outside the datapath. The schema records sqrt(dt/tau_ampa) as a derived constant for the enrolled default, avoiding a coarse generic square-root lookup in the OU scale. The generated Q32.32 Verilog compiles with Icarus and preserves the varied 32-update trace within:

• maximum physical-state error 2.5e-4;
• maximum rate error 0.30 Hz.

The rate envelope reflects the generic exprel lookup spacing. It is an explicit bounded co-simulation result, not exact transcendental equivalence.

Controlled benchmark

benchmarks/bench_wong_wang.py measures the same complete deterministic-sample batch through all five public dispatchers. It rejects an unpinned run, any unavailable runtime, a standalone Rust-safety failure, trace/final-state drift, or an invalid run size. The committed JSON binds raw timing samples, source hashes, exact loaded Rust/Go/Mojo binaries, tool versions, affinity, load, and every parity delta.

The record is one-logical-CPU, non-exclusive local regression evidence. It is not a production, cross-host, cross-framework, or hardware performance claim. The timing table is reproduced in Benchmark Comparison.

Boundaries

• Firing rates are continuous observables, not binary spikes.
• Gaussian samples cross the accelerator and RTL boundaries explicitly.
• The paper timestep is the maintained default; the author-code discrepancy is preserved in the evidence.
• Q32.32 co-simulation reaches H1 only.
• No synthesis, timing closure, formal equivalence, device resource, or PPA result is claimed.

Reproduction

The focused test invocation preloads the installed Rust extension before pytest adds repository support paths to the import search order.

Bash

rustc --edition 2021 --test \
  src/sc_neurocore/accel/rust/safety/wong_wang.rs \
  -o /tmp/wong_wang_tests
/tmp/wong_wang_tests

PYTHONPATH=src:. .venv/bin/python - <<'PY'
import sc_neurocore_engine
import pytest

raise SystemExit(
    pytest.main(
        [
            "-q",
            "tests/test_model_wong_wang.py",
            "tests/test_wong_wang_dynamics.py",
            "tests/test_reference_wong_wang.py",
            "tests/test_cosim_wong_wang.py",
            "tests/test_wong_wang_backends.py",
            "tests/test_bench_wong_wang.py",
        ]
    )
)
PY

taskset -c 11 env PYTHONPATH=src:. .venv/bin/python \
  benchmarks/bench_wong_wang.py \
  --json benchmarks/results/bench_wong_wang.json