One token.
Infinite connections.

Each layer computes z = W·a + b (a dense matrix–vector product; sparse layers store pruned weights as exact zeros under a mask), adds an identity skip + a[l−2] on residual layers, then applies its activation: ReLU, leaky ReLU (0.01), sigmoid, tanh, linear, or a numerically stable softmax on the output layer.

Training is mini-batch backpropagation: cross-entropy with softmax outputs, mean squared error otherwise. Optimisers: SGD with optional momentum, and Adam (β₁ 0.9, β₂ 0.999, ε 1e−8). Gradients are checked against central finite differences in the test suite.

Configs are canonically encoded (sorted keys, ECMAScript shortest round-trip numbers, −0 → 0, NaN/∞ rejected) and hashed with SHA-256. The PRNG neuron-prng/1 is sfc32 seeded from SHA-256 of the canonical config; normals use the Irwin–Hall sum of 12 uniforms so generation uses only exact integer and float operations.

config hash = H(engine, prng, config). artifact hash = H(schema, engine, config hash, every layer's W, b, mask, skip). result hash = H(engine, artifact hash, config hash, input, output). Engine: neuron-engine/1.0.0. Identical seed + config + engine version reproduce the same model hash.

Generation is bit-exact across conforming JavaScript engines. Inference uses Math.exp and Math.tanh, which the language does not require to be correctly rounded, so different engines or hardware may differ in the last bits. The reproduce action reports exact, within tolerance (|Δ| ≤ 1e-12 or relative ≤ 1e-9) or mismatch.

Recording an experiment appends its result, artifact and metadata hashes to the NEURON registry: an append-only log where every entry contains the hash of the entry before it and is signed with the server's ed25519 registry key. The database also refuses any update or delete of an entry.

Verified means the entry's hashes equal hashes recomputed from the stored experiment, the entry hash recomputes, the signature checks against the published key, and the entry links to its predecessor. It shows what was recorded, for which wallet and when, and that it hasn't been edited since. It does not prove the arithmetic was correct; anyone can re-run the pinned version to check that.

Models: ≤ 8 layers, ≤ 64 neurons per layer, ≤ 320 neurons, ≤ 12000 potential parameters, inputs within ±1000000. Training: ≤ 2000 epochs, ≤ 1000 samples, batch ≤ 256, in an isolated worker thread with a 256 MB heap cap and a wall-clock limit.

Generated models are not trained and carry no learned behaviour. Training on toy datasets improves only the measured loss and accuracy shown, on those datasets.

The API operator can see experiment inputs and holds the registry signing key. Because the log is hash-chained and its head is shown publicly on the Ledger, a rewrite of history is detectable by anyone who saved an earlier head or entry. Governance can only change whitelisted community-network parameters. Wallet sign-in uses single-use, domain-bound challenges; NEURON never asks for private keys and never moves funds.

Full details: docs/SECURITY.md and operations status.