About behavtaskatlas
A file-first atlas of sensory-guided decision-making tasks in behavioral neuroscience.
What this is
Task descriptions across behavioral neuroscience are usually written for a specific paper or lab, which makes them hard to compare. The atlas separates the operational variables of a task (stimulus, evidence schedule, choice geometry, response modality, feedback, training stage) from cognitive interpretations (perceptual decision-making, evidence accumulation, etc.), keeps both as YAML records, and links them to reproducible analyses against open public data.
Every record is a small file under version control. Every analysis is a script with pinned source data, recorded in provenance JSON. The catalog site you are reading is generated from those records.
Object model
- Task family — the reusable operational idea (e.g. random-dot motion, visual change detection).
- Protocol — one concrete variant of a family, with species, stimulus, choice, and timing fully specified.
- Dataset — a pointer to public or
private behavioral data, tagged with a
source_data_levelso reports rest-on-raw vs rest-on-figure-source-data is visible up front. - Vertical slice — one protocol bound to one dataset through a reproducible local pipeline (download → harmonize → analyze → report). Each slice produces canonical-trial CSVs, summary plots, and provenance.
- Implementation — task-control code or reference implementation, distinct from the dataset that contains the resulting behavior.
How records are validated
Every record is a Pydantic-validated YAML file. Cross-references between
families, protocols, datasets, and slices are checked, and controlled
vocabularies (modality, evidence type, choice type, response modality,
source data level, curation status) are enforced. The
release-check
command produces a release-readiness artifact that gates publishable JSON
exports.
How to contribute
Contributions go through GitHub pull requests against the YAML records.
Add or edit a record, run
bash scripts/ci.sh
locally, open a PR; Vercel will build a preview deploy automatically. The
repository includes a CLAUDE.md to make Claude Code sessions productive
without needing to read the full source.
Citation
Slice analysis artifacts are versioned as GitHub Releases tagged
slice-artifacts-<date>-<sha>.
Cite a specific release tag rather than the live URL when reproducibility
matters. A persistent DOI via Zenodo is planned for a future release.