Measured Behaviour Facet¶

Every model in the catalogue carries a measured behaviour facet: a small set of behavior_tags describing how it actually fired when probed across a sweep of constant-current drives. The tags are produced by simulation, never declared by hand, and the catalogue may only carry tags from a fixed vocabulary — so the facet describes observed behaviour rather than an author's claim.

The facet is independent of the citation tier. A model can be fully behaviour- tagged while remaining Tier 1 for want of a confirmable DOI, and a Tier 3 model that cannot be driven by a constant current carries no behaviour tags at all.

What the tags mean¶

Each tag has an exact, observable definition. A tag is asserted only when the behaviour was seen at some drive in the sweep.

Tag	Observed condition
`excitable`	fired at least one spike under some tested drive
`quiescent`	fired no spikes at any tested drive over the probe window
`tonic`	fired regularly (low ISI variability) under some sustained drive
`adapting`	spike-frequency adaptation — inter-spike intervals lengthening over time
`bursting`	a burst–pause structure (bimodal ISI) under some drive
`irregular`	moderate ISI variability under some drive
`chaotic`	high ISI variability under some drive
`phasic`	only one or two spikes despite sustained drive at some level
`rate-coded`	firing rate rose monotonically with drive (an increasing f–I curve)
`stochastic`	the spike train was not reproducible between two identical runs

excitable and quiescent are mutually exclusive, and a quiescent model never carries a firing-pattern tag.

How the measurement is taken¶

The probe (sc_neurocore.studio.behavior_probe) sweeps each model across a scale-robust ladder of constant currents:

Text Only

0.0, 1.0, 4.0, 16.0, 64.0, 256.0, 1024.0

The ladder spans several decades on purpose. The catalogue mixes unit conventions — physical millivolt/nanoamp models, normalised dimensionless maps, and integer fixed-point hardware models — so no single current is meaningful across all of them. A geometric sweep from rest to strong drive exposes each model's behavioural envelope regardless of its scale, where a single operating point would miss it (an adaptive-exponential model, for instance, only fires tonically under the strong-drive rail).

Each model is simulated over a 200 ms window. The firing pattern at each current is classified, and the tags are derived from the envelope of classifications plus the shape of the firing-rate curve.

Reproducibility and stochastic models¶

Every observation is taken twice from an identical seed. A firing-pattern tag is asserted only from an observation that reproduced exactly. A model whose spike train differs between the two runs — one with an internal, unseeded random source — is flagged stochastic and carries only the sign-robust excitability verdict; its fine pattern tags are withheld because they would flip from run to run. The derived tags are therefore reproducible by construction.

Models that cannot be driven¶

The probe is resilient per (model, current). A model that cannot be driven by a constant current — one whose step needs an extra synaptic input, or one with an internal stability guard that rejects the operating point — records the failure for that drive and still yields a profile. A model that fails at every drive is honestly left with no behaviour tags rather than a fabricated one.

In the Studio¶

The model browser shows a behaviour filter row built from the facet counts (GET /api/models/facets → behaviors). Selecting a tag restricts the list to models measured to show that behaviour. Unlike the on-demand Scan button — which classifies every model live at a single current — the behaviour facet is part of each descriptor and is available immediately, with no scan to run.

The facet is guarded by two cheap gates and one reproducibility gate:

Every descriptor's behavior_tags must be drawn from the controlled vocabulary (tests/test_behavior_taxonomy.py).
Every descriptor's committed tags must equal the recorded measurement in behavior_evidence.json (tests/test_behavior_tags.py), compared without re-running a simulation.
The probe's own reproducibility is proven on a sample of deterministic models (tests/test_behavior_probe.py).

The recorded measurement is regenerated with: