int-brain-lab/ibl-ai-agent/ibl-load

IBL Load

Use this skill when

• A scientific question requires neural or behavioral data loading.
• You need to choose between configured local datasets, aggregate/release tables, ONE objects, and session-level loaders.
• You need BWM loading policy.

References

• references/data_loading.md: short default loading policy.
• references/bwm_runtime_policy.md: canonical Brain Wide Map loading policy.
• references/bwm_fallback_preflight.md: required preflight when a BWM field is absent from local/aggregate tables.
• references/brain_regions_qc.md: spike sorting loaders, atlas mapping, region/QC handling, and plotting prep.
• references/bwm_ephys_spike_example.md: concrete local spike-shard example.
• references/repeated_site_pids.md: how to find repeated site PIDs, sometimes useful for exploratory analyses
• ../../docs/data_locations.md: machine-local data-root configuration.

Supplemental references:

• references/ibl_behavior_task.md
• references/bwm_release_scope.md
• references/bwm_paper_repo.md
• references/data_loading_live_one.md
• references/data_loading_fallback_examples.md

Workflow

1. For Brain Wide Map questions, follow references/bwm_runtime_policy.md.
2. For non-BWM questions, use references/data_loading.md.
3. Load the smallest surface that answers the question; restrict attributes, collections, and fanout.
4. Resolve large dataset roots from data_locations.local.yaml; do not assume reusable data live in this repo.
5. Prefer configured local datasets and cached files over remote-first examples when present and sufficient.
6. Use session-level loaders only when the requested metric cannot be obtained from local datasets, aggregate tables, or narrow ONE objects.
7. Record release freeze, local dataset path/version, data scope, and relevant cache/revision assumptions.

Quality gates

• Reject BWM plans that bypass an adequate local bwm_ephys or bwm_behavior dataset.
• Reject table choices with the wrong row grain when a correct-grain table exists.
• Reject generic fallback branches and high-fanout one.load_dataset(...) loops when a stable narrow local path is available.
• Fail fast on missing critical objects.
• Emit explicit warnings for optional missing modalities.