Registered data

Registration is a move, not a copy. md-openmm data-register verifies a finished directory, transactionally moves it into the managed storage root, and replaces the local directory with a symlink to the verified destination. Scripts that referred to the old path keep working; the bytes are in one place, under the contract.

What happens, in order

1. Verify. Every file is hashed, the machine records are read, completion is checked from those records rather than from prose in a log, and the manifest is built.
2. Move. The tree is staged into $MD_DATA at its canonical path and renamed into place.
3. Re-verify. Every digest is checked again at the destination.
4. Symlink. The original location becomes a relative symlink to the registered tree.

The source is preserved unless every check passes. A registration that fails leaves the local directory exactly as it was.

The canonical path

$MD_DATA/{year}/{project_name}/{data_name}/
$MD_DATA/common/{year}/{project_name}/{data_name}/     # role: common

year is the year the data were completed, not the year they were registered. There is no month segment. See project data for what a dataset is and what dataset.yaml records.

Registration is write-once

A registered dataset cannot be renamed, moved or added to. Two consequences worth knowing before you run the command:

• The data_name is permanent. Choose one that distinguishes this run from a later one on the same system — what changed, not just what it was.
• Anything a later tool needs must be inside the tree at registration time. A dataset that does not carry the inputs its own build record cites cannot be exported afterwards, and cannot be repaired in place.

Checking without writing

md-openmm data-register -idata <dir> ... --dry-run      # verify and print, write nothing
md-openmm data-register -idata <dir> ... --verify-only  # re-verify an already registered tree

--dry-run creates nothing, including the destination directory.