Skip to content

Worked examples

Each empirical example now carries a source-correspondence label. See the paper-figure register before interpreting “reanalysis,” “partial reproduction,” or “reproduction” as interchangeable claims.

Spontaneous events and long recordings

The spontaneous-transient sensitivity tutorial starts with known synthetic events, introduces an acquisition gap, and compares named threshold and local-baseline alternatives without claiming a tonic signal.

The tutorials are executable scientific narratives rather than isolated API snippets. Each begins with a question and records the experimental unit, preprocessing choices, event denominator, inferential target, and limitations.

Start with the public-data evidence atlas if you want to compare the scientific outputs and evidence boundaries before choosing a tutorial.

Public IBL feedback analysis

Import public IBL tables, inspect event coverage, and produce fingerprinted JSON and HTML evidence. This is the shortest complete example.

Open the IBL tutorial

Raw DANDI NWB to robustness report

Run a six-animal reward analysis across eight declared preprocessing universes. This example demonstrates NWB ingestion, animal-level inference, explicit incompatibility, failure retention, and report verification.

Open the DANDI tutorial

Event-kernel encoding simulation

Recover overlapping cue and reward kernels while controlling a continuous motion covariate and holding out complete animals. This is the ground-truth implementation-validation companion to the public-data example below.

Open the event-kernel simulation

Event-kernel model multiverse

Compare named cue, reward and motion design specifications under one fixed animal-held-out validation policy. Score deltas are only reported when models use the exact same retained timestamps, and failed designs remain in the ledger.

Open the model-multiverse method

Previous-outcome event kernels

Fit an average cue kernel together with an explicitly coded previous-outcome modulation. The simulation demonstrates recovery, session-boundary resets, and the separation between within-session neural encoding and Unspool's longitudinal behavior models.

Open the event-history tutorial

Variable-duration behavior kernels

Keep physical bout boundaries while jointly modeling onset, duration modulation, and normalized within-bout progress. Outside-bout samples remain in the continuous recording denominator.

Open the variable-duration tutorial

Public DANDI event-kernel reanalysis

Fit joint active-poke and reward-increment kernels to DMS and DLS recordings from six checksum-pinned public animals. The example foregrounds weak held-out prediction and a boundary-selected ridge penalty rather than hiding them behind pooled coefficient shapes.

Open the public event-kernel reanalysis

Public IBL longitudinal neural–behavioral forecast

Compose FiberPhotometry's checksum-verified session neural summaries with Unspool's cohort-forward behavioral validation. The retained result shows that the previous session's coarse DMS feedback contrast does not improve prediction in the declared future session.

Open the cross-package longitudinal tutorial

Pose and behavior-tool interoperability

Compose DeepLabCut or SLEAP pose confidence, Keypoint-MoSeq bouts, and BORIS point/state annotations with photometry covariates and events. The tutorial then passes declared neural summaries to Unspool without duplicating longitudinal behavior models.

Open the ecosystem interoperability tutorial

Literature reproductions to add

The maintained paper-figure register names exact source panels, current departures, and acceptance criteria. This replaces an open-ended list of papers with testable reproduction targets.