neuroatlas transform artifacts v1

Publishes the qualified MNI152NLin6Asym ↔ MNI152NLin2009cAsym ANTs image-transform pair, evaluated on 1 mm and 2 mm target grids. Large transforms are separate release assets; they are not bundled with the R package.

Both independent builds passed all eight numerical cells under the prospectively frozen policy, and independent agent review inspected all 24 QA panels. Repeat-point differences are zero; maximum own round-trip error is 0.078176 mm against the 0.5 mm limit. Independent-engine point error is below 0.000474 mm. All evaluated Jacobians are finite and positive within the brain masks.

The evidence archive retains the original measurements, image round-trip diagnostics, native ANTs references, hash-bound reviews, input/software provenance, and failed earlier attempts. At 2 mm, minimum label round-trip Dice is 0.796602; inverse 2 mm minimum HOCPA concordance is 0.577864. These mapped labels are diagnostic, not independent anatomical truth. This qualification does not establish individual-subject accuracy or global Jacobian positivity.

Download and extract qualification-evidence.tar.gz, then open report.html. Use SHA256SUMS to check the assets. Each H5 is approximately 86 MiB. See LICENSES.md for template and evidence distribution terms, including FSL commercial-use restrictions. The R source remains MIT licensed.

The immutable manifest is transform-artifacts-v1.json. Public package integration is verified before enabling the registry rows in neuroatlas. A changed artifact, recipe, input, or qualification policy requires a new release.