Scientific contracts

These contracts describe snnlab's implementation. The equations below restate its timing and input conventions; they are not a new neuron model.

Units and tensor axesCopy Anchor Link

Projection weights use microsiemens (uS), matching leak_us. Other projection-unit labels are rejected rather than silently rescaled. Some historical pre-0.2.0 authoring bundles carried an incorrect nS label despite executing values as uS; inspect provenance before interpreting old artifacts.

Time-varying inputs use (time, batch, channels). Spike values are binary. Masks are boolean or zero/one. Voltage parameters use millivolts, and graph timesteps use milliseconds.

Physical durationCopy Anchor Link

For $N$ simulation steps and timestep $\Delta t_{\mathrm{ms}}$ measured in milliseconds, the duration $T_{\mathrm{ms}}$ in milliseconds is

$$T_{\mathrm{ms}} = N\,\Delta t_{\mathrm{ms}}.$$

The default Network timestep is 0.1 ms. Physical durations and integer step counts are not interchangeable. The PING component declares refractory periods as explicit E/I step counts; changing timestep therefore changes their physical duration.

Discretised Poisson inputCopy Anchor Link

Generated homogeneous Poisson inputs use a Bernoulli draw per channel and timestep with probability $p$:

$$p = r_{\mathrm{Hz}}\frac{\Delta t_{\mathrm{ms}}}{1000}.$$

Here $r_{\mathrm{Hz}}$ is the requested rate in spikes per second; $\Delta t_{\mathrm{ms}}$ is the timestep in milliseconds, and 1000 converts milliseconds to seconds. $p$ is dimensionless. Rates that make $p>1$ are rejected. This convention permits at most one emitted input spike per channel per timestep.

Delays and causalityCopy Anchor Link

Positive projection delays must be exact integer multiples of the graph timestep. Zero-delay feedforward connections follow deterministic topological order. Recurrent and feedback paths remain causal: zero additional delay means one simulation step. Zero-delay cycles and unsupported structures fail before simulation.

Graph, recipe, and protocolCopy Anchor Link

Keep physical dataset and checkpoint paths outside the authored graph. Authenticate artifacts before reuse. Matching seeds alone do not prove equivalence across devices or software versions.