Stochastic & Neuromorphic Computing  /  Reference  /  Source fidelity

Source fidelity
every model traced to its paper

A neuron model is only as trustworthy as its provenance. In SC-NeuroCore, each bundled model is independently re-derived from its primary source and checked feature by feature against a committed contract with explicit tolerances — anchored to a paper by DOI, cross-checked across five language backends, and pinned so it cannot silently drift. This page shows that machinery, and lets you look up any model's source yourself.

The provenance table

Every model below carries a DOI to the paper it is derived from. Type a name, an author, or a DOI to filter. This is generated from the bundled model schemas themselves — the same objects the library runs and validates — so the table is the ground truth, not a hand-kept list.

No model matches that filter.
ModelSource (DOI)TypeStateParams

51 DOI-anchored models shown; a handful of models with in-progress schema detection are omitted. The full literature-transcribed catalogue lives in the neuron atlas, and the live dynamics on the neuron-models page.

How a model is validated

Each model has a committed reference entry: the schema, a deterministic protocol, the source provenance, the expected scalar features (spike count, rate, sub-threshold shape, event timing) and a per-feature tolerance. The production validator runs the exact same UniversalNeuron that public workflows use, and reports any feature-level mismatch without falling back to another trace — a model either meets its contract or it fails, visibly.

Deeper: independent re-derivation, not self-comparison
The reference is not the model checking itself. Each contract is an independent re-derivation of the published equations — an analytic linear recurrence where one exists, forward explicit-Euler for sub-threshold dynamics, or a macro-step RK4 (e.g. 100 inner dt=0.01 ms sub-steps per 1 ms macro step, with a macro-boundary threshold crossing) for stiff conductance models like Hodgkin–Huxley and Connor–Stevens; discrete maps are iterated literally from the paper's equations. The implementation conventions (reset, wrap, event definition) are stated separately from the DOI-sourced continuous equations, so the science and the numerics are auditable apart.
Deterministic and statistical references

Deterministic models get scalar-feature contracts. Stochastic ones — an escape-rate neuron, a Poisson source — cannot be pinned to a single trace, so they carry separate seeded statistical references with exhaustive distribution and event-hash checks over a full period, rather than being coerced into the deterministic schema.

deterministic: feature ∈ [expected − tol, expected + tol] · stochastic: seeded distribution + event-hash over full period
Deeper: why not just diff two floating-point traces
A raw point-by-point trace diff is brittle — a one-step timing shift or a last-bit rounding difference fails a comparison that a scientist would call correct. Feature contracts with tolerances encode what actually matters about a model's dynamics (does it burst, at roughly this rate, with this sub-threshold shape) and make the pass/fail meaningful. For stochastic units, a single sample proves nothing, so the reference is the whole seeded distribution and a hash of the event sequence.
Parity across languages and silicon

The same model is not one implementation. Its dynamics are checked bit-consistently across the polyglot backends — Python reference, and the Rust, Julia, Go and Mojo kernels — and again against the generated fixed-point Verilog in co-simulation. A committed reference receipt records the outcome, so a model that passes in Python but drifts in a fast kernel is caught, not shipped.

Copyright, done properly

Provenance is not the same as redistribution. The copyright-bound publisher PDFs are never committed — they are kept out of the repository entirely. What is tracked is only SC-NeuroCore's own attributed parameter tables and a DOI index, so anyone can obtain a paper from its DOI and check the transcription against the note. Sources carry their licenses (many CC BY; several © publisher) in that index.

What this proves — and what it does not

Honesty is the point of the exercise, so the boundary is explicit. The corpus validates each model against its committed deterministic feature contract (and, for stochastic units, its seeded statistical reference). It does not claim replay against external simulators — NEST, Brian2 or NEURON — nor against published figures; those are separate validation surfaces, kept separate rather than quietly conflated. A green corpus means "meets its own re-derived, DOI-anchored contract", which is a strong and checkable claim, not a universal one.