Validated NWB conversion

Supported streams

The executable workflow covers two explicit input streams in one session:

Other acquisition readers require their own format-specific tests. In particular, this helper does not decode SpikeGLX, Open Ephys, multichannel TIFF, volumetric TIFF, compressed video, or pixel-to-world calibration. Do not rename an arbitrary numeric table to a supported stream.

Install the tested environment

uv venv --python 3.12 nwb-env
uv pip install --python nwb-env/bin/python 'neuroconv[tiff]==0.10.2' pynwb==4.2.0 \
  nwbinspector==0.7.2 roiextractors==0.10.0 tifffile==2026.9.20 \
  zarr==2.18.7 hdmf-zarr==0.11.3

Keep both Zarr pins even for an HDF5-only conversion: NeuroConv 0.10.2 imports its backend configuration modules at startup, and the tested unconstrained Zarr 3.4.0 installation failed on zarr.codec_registry. The pinned environment ran the real conversion, PyNWB validation and Inspector successfully on macOS ARM64. The dependency resolver supplies NumPy and HDF5 support. These are compatibility pins, not claims that Zarr 2 and hdmf-zarr 0.11.3 are the latest releases. Current interface checks and the tested dependency exception are recorded in references/upstream-review.md.

Workflow

1. Inventory the actual inputs and acquisition metadata. Identify image plane/channel, optical settings, subject/session identifiers, timezone, behavior coordinate system, units and the timestamp clock for every stream. Preserve originals. Do not replace missing metadata with plausible defaults from a sample config.
2. Copy assets/session-template.json beside the raw data and replace the explicitly synthetic values. Paths resolve from that JSON file. Read references/input-contract.md for the exact CSV and metadata contract and the pulse-pair variant. TIFF pixels are retained as acquired; a raw arbitrary-unit intensity does not become a photon count merely by changing its unit label.
3. Establish the common timebase from acquisition evidence. Frame timestamps must already be reference-clock seconds since the timezone-aware session start. For position, provide either a documented shared clock or matched synchronization pulses. The helper fits one affine clock transform, checks its residual against a specified tolerance, and refuses extrapolation beyond the pulse range. It never estimates synchronization from coincident-looking neural/behavioral signals. Clock resets or nonlinear drift require an explicitly validated piecewise mapping.
4. Execute the converter. Inputs must have finite, strictly increasing timestamps and matching image/timestamp counts. Explicitly declare one channel and one plane; known TIFF channel/plane metadata must agree. Grayscale pages alone cannot exclude undocumented interleaving. The acquisition samples stay intact; only coordinate units and, when evidenced, behavior timestamps are transformed.
5. Read the .validation.json alongside the NWB file. Schema compliance, Inspector findings and data equality answer different questions. The script exits with an error for schema failures and flags critical Inspector findings for review in the report. Review all findings in context; successful validation cannot establish that anatomical labels, pulse pairing or calibration supplied by the user are correct.
6. Deliver the NWB, validation JSON, original conversion config and an explanation of remaining metadata gaps or Inspector findings. No upload or archive submission is part of this workflow.

Execute

Run the following from the skill directory, with paths to the actual analysis files:

nwb-env/bin/python scripts/convert_session.py /path/to/session.json --output /path/to/session.nwb

nwb-env must point to the environment created above; the absolute example input paths are

…