McCulloch-Pitts source-to-silicon fidelity evidence

This page records the primary-source rule, independent truth corpus, five-language parity, stateless schema execution, Python-to-Verilog co-simulation, bounded formal job, and source-bound benchmark used to promote McCullochPittsNeuron to the polyglot-complete catalogue.

Primary source binding

The scientific source is McCulloch and Pitts (1943), A Logical Calculus of the Ideas Immanent in Nervous Activity, doi:10.1007/BF02478259. The source rule provides three relevant single-cell invariants:

1. nervous activity is all-or-none;
2. excitation requires a fixed number of active excitatory afferents within one synaptic delay;
3. any active inhibitory afferent prevents excitation absolutely.

The maintained model exposes these invariants directly. A real-valued weighted sum with negative inhibitory weights is a later abstraction and is not used as a substitute. The one-synaptic-delay statement defines network scheduling; it does not justify invented membrane or delay state inside the neuron.

Independent reference

tests/test_reference_mcculloch_pitts.py implements the logical rule without importing production model code. It canonicalizes the eight primary truth rows as compact JSON and pins SHA-256:

Text Only

2aebd2a5ed6ea8a9409b7452d603441d91e2c8b610da8446668452492ee73db4

The corpus covers:

theta	Excitatory count	Inhibition	Output
1	0	false	0
1	1	false	1
1	0	true	0
1	1	true	0
2	0	false	0
2	1	false	0
2	2	false	1
2	2	true	0

The committed artifact src/sc_neurocore/neurons/reference_trace_data/mcculloch_pitts_1943_truth_table.json binds the DOI, official paper page, primary reprint, equation statement, schema, standard reference runner and zero-tolerance features. The hand model, TOML schema and JSON schema reproduce the enrolled output exactly.

Executable evidence matrix

Surface	Executed contract	Result
Python	strict int32 count/threshold, Boolean inhibition, absolute veto, varying batch	source rows and boundaries exact
Rust engine	stateless same-name PyO3 class and full batch function	trace and count exact
Rust safety	standalone module compiled with rustc --test	9/9 tests pass
Julia	complete integer count/flag batch	trace and count exact
Go	validated service and generated C-shared ABI	trace and count exact
Mojo	validation pass followed by atomic output pass	trace and count exact
TOML/JSON DSL	empty state set and strict level threshold	source rows exact
Registered RTL	state-owning timing shell over signed Q32.0 comparator	encoded truth vector exact
Folded RTL	combinational signed Q32.0 datapath	encoded truth vector exact
SymbiYosys	depth-4 Z3 reset-spike safety	PASS

The batch PyO3 implementation lives in engine/src/bindings/mcculloch_pitts.rs and is registered through the existing pyo3_neurons registry. Relative to the accepted parent, Model 33 leaves engine/src/lib.rs unchanged at 7,116 lines and 212 #[pyfunction] entries. An architecture regression test prevents this binding from returning to the crate root.

Boundary tests separately prove maximum signed count acceptance, threshold equality, zero excitation, complete-veto dominance, strict Boolean flags, zero-row batches, malformed native-output rejection, explicit-backend failure, and atomic Go/Mojo rejection before destination writes.

Python-to-Verilog contract

The paired schemas contain no state equations. They encode only

Text Only

spike = I >= theta

on a signed Q32.0 port. Public encode_hardware_input maps an uninhibited non-negative count to itself and maps any active inhibitory afferent to -1. Since theta >= 1, the sentinel can never satisfy the threshold. The enrolled Icarus test drives zero, subthreshold, equality, suprathreshold, maximum-int32 and inhibited rows through both production forms; Python, registered RTL and folded RTL return the identical binary vector.

The registered module holds only its output timing register. It does not imply biophysical state. The folded module is purely combinational, allowing a population engine to own network scheduling explicitly.

Controlled five-backend benchmark

benchmarks/bench_model_mcculloch_pitts.py performs a 1,000-row warm-up and seven 200,000-row calls through each public dispatcher. Counts cycle from zero through fifteen, the last count is maximum signed int32, and every eleventh row activates inhibition. It fails if a backend is missing, an event differs, a count differs, the process is unpinned without acknowledgement, or the standalone Rust safety module fails.

The committed artifact records raw samples, exact input and output hashes, source hashes, loaded Rust/Go/Mojo binary hashes and sizes, runtime versions, affinity, governor and load averages. The run uses one logical CPU without claiming exclusive isolation. Its timing fields are local regression evidence, not a production speed or hardware claim.

All five lanes emit 102,273 events and the identical binary trace SHA-256 52a05b62f801b9a9856ccac9f6d79f2821d564239b85fd06d454d1d44e28aee4. End-to-end public-dispatch medians are 234.741/306.158/328.712/625.298/821.117 ms for Rust/Go/Python/Mojo/Julia, including common Python input validation. The CPU-4 powersave run began at load averages 59.52/52.70/47.26 and makes no exclusive-isolation or portable ranking claim.

Descriptor and formal boundary

McCullochPittsNeuron.toml records the complete authorship, DOI, parameter and stateless contracts, five exact backends, reference digest, science S5 truth evidence and silicon H1 evidence.

The catalogue emitter produces sc_mccullochpittsneuron.v, its port-only formal harness and sc_mccullochpittsneuron.sby. The depth-4 BMC proves the declared reset-spike safety property. It does not prove unbounded equivalence, absence of overflow for arbitrary invalid external encodings, synthesis timing, placement, power or physical-device behavior.

Scope boundary

• The source logical rule is implemented. Perceptron learning, differentiable surrogates and arbitrary real weights are separate models.
• Network topology, synaptic propagation and the one-delay scheduler remain caller responsibilities.
• -1 is an explicit transport sentinel, not a claim that every negative hardware word has a biological interpretation.
• The model is a deterministic simulation and research RTL primitive, not a medical, safety-certified or deployed neuromorphic device.

Reproduction

Bash

PYTHONPATH=bridge:src:. .venv/bin/python -m pytest -q \
  tests/test_model_mcculloch_pitts_logic.py \
  tests/test_model_mcculloch_pitts_validation.py \
  tests/test_model_mcculloch_pitts_hardware_encoding.py \
  tests/test_model_mcculloch_pitts_batch_dispatch.py \
  tests/test_model_mcculloch_pitts_network.py \
  tests/test_mcculloch_pitts_schema_dsl.py \
  tests/test_reference_mcculloch_pitts.py \
  tests/test_mcculloch_pitts_backend_loading.py \
  tests/test_mcculloch_pitts_backends.py \
  tests/test_cosim_mcculloch_pitts.py \
  tests/test_bench_mcculloch_pitts.py

taskset -c 4 env PYTHONPATH=bridge:src:. .venv/bin/python \
  benchmarks/bench_model_mcculloch_pitts.py \
  --json benchmarks/results/local_python_2026-07-14_mcculloch_pitts.json

cd hdl/formal/catalogue
sby -f sc_mccullochpittsneuron.sby