neuralos-snn

no_std, i16 fixed-point spiking neural networks for edge and RISC-V silicon — LIF neurons, pairwise STDP, CSR synapses, ternary weight codecs, and an AVX2 batch kernel.

Published on crates.io as 0.1.0-alpha.6 (2026-09-11; AGPL-3.0-or-later).

What this crate is

A spiking-neural-network substrate: the neuron/synapse/network core you can run bare-metal. Integer-only hot path (i16 millivolts, i16 microamps, u32 microseconds) — no floating point, no allocator in the core, no_std by default. Designed for FPU-less edge targets (ESP32-C3, HiFive, QEMU riscv64gc) and validated against the IEEE 2025 "Full-Integer SNN Inference with RISC-V ISA" design axis.

Modules

The voltage grid story

delta_v = dt_over_tau · (leak + R·I/1000) / 1000 truncates to whole quanta. On the default mV grid a steady sub-threshold current inside the ~200 μA dead zone moves the membrane exactly zero — recorded, tested, and the reason VoltageResolution::CentiMillivolt exists (100× finer dead zone, same i16, bit-identical arithmetic shape). The mV default keeps every historically recorded result bit-exact.

Features

• std (default) — enables the network orchestration module
• simd — implies std, x86_64-only AVX2 batch kernel

Without std the crate builds no_std (neurons, synapses, trit, bridge, kernel, nir) — the embedded posture CI enforces.

Usage sketch

use neuralos_snn::{SpikingNeuralNetwork, NetworkTopology};

let mut net = SpikingNeuralNetwork::new_with_voltage_resolution(
    128, 1_000, NetworkTopology::Balanced { excitatory_ratio: 0.8 },
    Default::default(),
)?;
net.build_topology()?;
loop {
    let spikes = net.step(&inputs)?;   // decay → integrate → clear → propagate
    // spikes: Vec<Spike>; plasticity applies pairwise STDP when enabled
}

Anti-scope

No I/O, no persistence, no UI, no drivers, no "OS", no crypto, no LLM runtime — the library is a library. (The research runtime that proved a ternary SNN↔LLM bridge on this substrate lives in the workspace's neuralos-rt, publish = false.)

Since alpha.2

• F1 — adaptation-decay contract pinned: unit tests pin the exact −1/step decay with floor 0 and the +2/spike jump; a 6,000-step live test proves a driven net stays firing and adaptation equilibrates. Leak convergence pinned too (dt = τ lands on rest exactly, both directions).
• The coupling knob — synaptic_input_divisor (new API): the recurrent pulse is weight / divisor μA. Default 10 = the historical pulse byte-for-byte; 0 rejected; pinned by default + doubling tests.
• network.rs split into csr.rs + stats.rs — every published path unchanged, CSR build/equivalence now pinned by dedicated tests (unsorted insertion, external adds, reverse-CSR incoming, plasticity weight sync).
• F5a — the DECORATIVE-in-orchestration machinery removed (alpha semver window): Synapse's transmission/eligibility state (delay_us, conductance, transmission_buffer, eligibility_trace, recent_activity, synapse-side last_spike_time_us) and its dead methods (transmit, receive_spike, is_active, set_delay_us, reset), plus LIFNeuron::tau_synapse_us. Orchestration never read any of it — proven output-neutral by byte-exact re-pins of the bridge examples (gate verdict, export sha, 13/13 null patches).
• Pub-API census: five module-internal fns de-pubbed (LIFNeuron::new_with_type, set_voltage_resolution, the three i2_s layout helpers); introspection accessors and builders stay pub — see the alpha.3 audit record in the repo's ISA.md.
• SIMD gate runs in CI; the batch kernel is documented mV-grid-only.

Since alpha.6 (the alpha.7 notes)

• Both divisions by 1000 in integrate_and_fire narrow to i32 when the value fits (private div_1000): the same truncating division of the same value, a multiply-high by the constant on a 32-bit core instead of a call to a software 64-bit division; a value that does not fit takes the i64 division as before, out of line. Exact by construction, pinned by a proptest against x / 1000 in i64 (half its draws in i32, half in all of i64) and a unit test on both sides of the guard; the two i128 exactness proofs of integrate_and_fire are unchanged and green.
• dt_over_tau computes in u32 when dt_us <= u32::MAX / 1000 (4,294,967 µs): one hardware divide on a 32-bit core; above it, the u64 formula as before. The same results over the whole u32 × u32 domain, pinned the same way.
• The spike ring indexes its two stores with a mask (& (MAX_SPIKE_HISTORY - 1), the power of two asserted at compile time): the same slots, and a bound the optimizer can see in the index. On the ESP32-C3 the bounds check on the bare index was what kept the neuron out of registers.
• #[inline] on LIFNeuron::integrate_and_fire, a hint: without it one of the firmware's call sites stepped the neuron out of line.
• On the ESP32-C3 (rv32imc, 160 MHz), same board and same firmware source, the firmware now times two bursts. The neuron as a network holds it (in memory, dt at run time) costs 1,579 ns per step, against 3,878 on alpha.6; the burst the compiler can see through costs 563, against 2,863. Every behavior figure is identical: 147 spikes in the 10,000-step burst, the first at step 55, the same checksum of the spike steps; the real-time loop's first spike still at step 56. No 64-bit software division is left in the step. The 2,842 of the alpha.6 notes below is the second burst's shape on alpha.6, measured before the checksum code existed; a network's run-time dt cost more, 3,878. Mechanism, listings and pins: evidence/esp32c3-bringup/README.md § Third entry (ISA round 27).
• On x86-64 the host spike-path bench is about 2 to 2.5 ns per step slower (+22 to 26 %), bisected across the round's commits and recorded, not fixed (the principal's ruling): the chip is the target. evidence/esp32c3-bringup/README.md § Host bench, round 27.

Since alpha.5 (the alpha.6 notes)

• Spike history is a [u32; MAX_SPIKE_HISTORY] ring (private SpikeRing in lif_neuron.rs): O(1) push, the oldest entry overwritten once full, iteration oldest to newest. It replaces heapless::Vec, whose remove(0) shifted the whole buffer on every spike past the 64th. Proven by a differential proptest against the pre-ring implementation kept verbatim as the oracle (same spike times, same order, same three consumers), a proptest on the ring alone, and two unit pins (Debug shows live entries only; indexing past len panics like a slice). No public API moved.
• No runtime dependencies. heapless is a dev-dependency now (the oracle above); cargo tree -e normal prints the crate alone. The crate depends on core.
• rust-version = "1.92" declared, inherited from the workspace; it is the toolchain pin, the only MSRV anyone has verified.
• isi_stats_us sums in u128 (test-gated, zero non-test callers): with u64 accumulators, three intervals above 2^32/√3 (about 2.48e9 µs, reachable through the non-monotonic filter) overflowed the sum of squares. Found by the differential proptest on 2026-09-10; that test now runs on arbitrary u32 times, and a unit test pins three maximal intervals on both sides.
• On the ESP32-C3 (rv32imc, 160 MHz) the timed burst costs 2,842 ns per neuron step with this version, against 1,501 ns measured on alpha.5 with the same firmware source, every behavior figure identical (147 spikes in 10,000 steps, first spike at step 56, 14.7 spikes/s). Read from the two flashed binaries: the alpha.5 burst loop was one the compiler happened to keep in registers, with one 64-bit software division per step; alpha.6's keeps the neuron in memory and pays two, plus a hardware divu, which is what a Vec<LIFNeuron> network pays on either version. The library did not get slower in the general case; the honest number is the higher one. Mechanism, listings and pins: evidence/esp32c3-bringup/README.md § The mechanism (ISA round-23); narrowing those divisions, bit-exact, is the next brief.

Since alpha.4 (the alpha.5 notes)

• General graph assembly — NirImport::build_network: any reference-emitted Input/Linear/LIF/Output graph assembles onto a real SpikingNeuralNetwork and fires. Every LIF population becomes neurons (its own quantized params); every LIF→LIF edge becomes an EDGE_PULSE_QUANTA synapse pair (the ratified D1 contract: 200 → a 20 μA pulse at the default divisor 10, +10 centi-quanta exactly one step later — exact-pinned on both grids, the mV grid dead 10× over, which is why recurrent graphs reject mV options BY NAME with the copy-pasteable re-import remedy); the Linear DAG folds symbolically at setup and quantizes ONCE per (drive Linear, root Input) stage (D2 fusion — no hop-by-hop i16 composition), merged saturating into the global per-step current vector (D5 multi-Input). Plasticity frozen at assembly (NIR has no plasticity term). New public surface: build_network, NirGraphEncoder, NirAssemblyReport, LinearFusedRecord, EDGE_PULSE_QUANTA, and NirError::EdgeShapeMismatch (per-edge reference type-check parity). Every out-of-slice shape is a NAMED rejection (readout, direct drive, encoder-only, self-loop, Output-as-source, Linear cycles, pass-through, empty/no-Input/no-Output, the D7 population bound); Input-unreachable structure assembles with structural UndrivenPopulation notes — silence documented, never silent. The frozen chain is proven bit-exact through both builders (nir_assembly_gate, 6/6, evidence of record).
• Honest-claim language of record (assembly): structure, per-edge type shapes, and quantization records are exact w.r.t. the pinned document; dynamics are named substrate conventions (EDGE_PULSE_QUANTA, divisor, grid, one-step delay, frozen plasticity) pinned by exact-value tests; the reference defines no execution semantics at this sha.
• The banked consolidation breaks ride here (2026-08-22): the NeuronBuilder/SynapseBuilder types and the setter surface deleted, the test-only introspection quartet relocated under #[cfg(test)], LIFNeuron::spikes() deleted (zero callers), Synapse::normalized_weight test-relocated — API breaks with no known consumers.

Since alpha.3 (the alpha.4 notes — PUBLISHED 2026-08-22)

• NIR slice 1 shipped in alpha.3 (missing from the notes above): neuralos_snn::nir — JSON import/export of Input/Linear/LIF/Output graphs, schema pinned to neuromorphs/NIR@7883c3c, reference-emitted fixtures + nir_format_gate, explicit per-node quantization records.
• Fresh-eyes review fixes (findings R1-R8): skip_value depth-capped at 64 (adversarial nesting is a loud Json error, not a stack overflow); trailing content after the root rejected (Python json.loads parity); denormal-absmax weights (e.g. 5e-324) are a loud BadNumber — the underflowed scale = 0 would have zeroed exports silently and broken idempotence; round_half_away rewritten as truncate-compare (the classic add-±0.5 idiom misrounds values 1 ulp below a half — reachable via r in MΩ: 499999.99999999994 Ω imported as 1 MΩ); ChainEncoder::encode accumulates in i64 (i32 silently wrapped negative at cols ≥ 3 with full-scale weights); scan rejects 1-D/empty/3-D weight arrays up front; export rejects dangling edge indices (no "?" placeholders).
• Structured-entry seam (post-review): quantize_linear and quantize_lif are public — callers holding materialized f64 values (HDF5 import, builders) quantize without a JSON document, the same contract and errors, arena placement included. The 4× arena-scratch trick is gone: import stages source weights in a typed NirBuffers::scratch f64 slice; arena and scratch each hold the weight-cell count exactly. Breaking buffer-API change, alpha.4-bound.
• Per-neuron LIF populations + NirBuilder: a LIF node is the reference's population — per-neuron param arrays quantize to per-neuron records in a lifs buffer (NirLifPopulation views); assembly enforces Linear rows == population size (reference type-check parity), one neuron per row with its own params. NirBuilder (std) assembles graphs in memory over the quantizer seam; export renders per-neuron arrays; the printable-ASCII gate is a JSON-container property only (new NonAsciiNodeName at export, never on the typed surface). Reference-emitted chain_population.json fixture pins the expansion end-to-end (imports → assembles → fires → re-imports state-identical).
• Honesty note: findings R1-R5 were present in the published alpha.3 binary. No consumers are known (the module shipped within the day); all are fixed here, ahead of any consumer, targeted for alpha.4.
• The HDF5 .nir container ships workspace-side, in neuralos-rt behind its hdf5 feature (vendored static HDF5, pre-read filter census, the reference's own fixtures + nir_hdf5_gate) — the seam it feeds is the structured entry above.

Status

0.1.0-alpha.6 is live on crates.io (published 2026-09-11): the [u32; N] spike ring, no runtime dependencies, rust-version declared, isi_stats_us in u128, and the ESP32-C3 step cost read honestly (see "Since alpha.5"). Alpha.5 (2026-08-22T13:59Z from 103fa59, registry-verified): the general four-kind graph assembly, EDGE_PULSE_QUANTA, the assembly gates, the R17 consolidation breaks and the STDP dt-overflow fix (see "Since alpha.4"). Alpha.4, the same day: NIR structured entry (pub quantizers, NirBuilder, per-neuron populations) + the R9 review fixes. The alpha.3 record: the adaptation-decay contract pinned by unit + live tests (equilibrates, never silences the net); synaptic_input_divisor — the coupling knob, new public API (default 10 = the historical weight/10 pulse; 0 rejected); network.rs split into csr.rs + stats.rs with every published path unchanged; the simd batch kernel doc'd mV-grid-only.

The tree is ahead, alpha.7 pending (the divisions and the ring mask; see "Since alpha.6"). Test counts live in the CI log and the repo's AGENTS.md § Commands, not here; the API may still move within alpha semver.

NIR itself: Pedersen et al., Nature Communications 15, 4962 (2024), DOI 10.1038/s41467-024-52259-9 — this crate's nir module speaks the schema pinned to neuromorphs/NIR@7883c3c.

License

AGPL-3.0-or-later — see the workspace root LICENSE.