Source audit
The audit pins nine published DANDI versions, inventories 260 immutable NWB assets, and maps all 118 original WormID assets plus the exact 36 tracking acquisitions (EY 21, KK 9, SK1 4, SK2 2). SF's 38 records remain weakly labeled, and NP/HL remain structural. The original split file has two repeated structural rows. Raw image shape is checked on acquired or range-read pilot assets, not inferred solely from publications.
The ZephIR ZIP central directory was range-read before acquisition. Only small worm annotation files and one 125 MB compressed image member were acquired; the mouse recording and all macOS metadata are excluded. Per-member extraction checks ZIP CRC, while the full 14.9 GB archive MD5 is only an upstream reference and is not verified locally. Targettrack's 164 MB volumetric archive and ASCENT CSV files pass source MD5 checks. Derived/selected image assets have their own SHA256 checksums; selective NWB/HDF5 reads do not verify the full source file checksum.
Shared DANDI subject identifiers produce potential overlap groups. DANDI 001623 overlaps identifiers from SK1, so independent specimen counts cannot be inferred by summing registry rows. NeRVE-in-ZephIR is a derivative of an existing acquisition. Skuhersky and NeuroPAL canonical CSVs appear identical in published Git trees; they are references only and not counted as two anatomical datasets.
Known blocks: CeNDeR OSF native storage lists no files, linked Google Drive API returns 404, and OSF data license is unstated. BrainAlignNet Dataverse API did not return usable metadata. OpenWorm movement hostname fails DNS. NeRVE original IEEE page resolves but its full recording inventory and download terms are not completed. Reusable population atlas licensing and an eigenworm basis remain unresolved. None of these blocked sources is silently replaced with synthetic data.
A total source storage inventory is an upper bound that includes duplicates. The original SF corpus alone exceeds 1 TB. This alpha release selectively acquires small source files and range reads and publishes a compact subset (exact hosted bytes in the assembly manifest; total HTTP transfer bytes were not measured), avoiding duplicated multi-terabyte movies. Extended full acquisition and hosting require a deduplicated transfer budget. No scheduler jobs or immutable experiment worktrees are modified.