Repository: plastic-labs/registeel
Repositories use Registeel to record job provenance. Before work begins, the repository creates a manifest containing the resolved configuration, repository state, Python environment, and selected input-file hashes. It stores the manifest and its ID in the job record, linking later artifacts to those inputs.
A manifest identifies the recorded inputs behind an artifact. Comparing a new manifest with the selected baseline reveals unintended changes before committing expensive compute.
Manifest contents
| Area | Recorded information |
|---|---|
| Job | Manifest type, schema version, creation time, and job ID |
| Configuration | Resolved configuration after key-based secret redaction |
| Code | Revision, branch, remote, and uncommitted changes |
| Python environment | Selected package versions and a hash of all installed distributions |
| Input files | Path, size, and SHA-256 content hash for each selected file |
Each repository defines its configuration schema and selects the packages and input files to record. Registeel does not infer them from configuration values. The manifest describes the job at startup; it does not contain datasets, metrics, checkpoints, models, or other artifacts created afterward.
Manifest identity
The manifest ID is the SHA-256 hash of its deterministic JSON representation. The local store verifies the hash on write and read, detecting edits and corruption. The ID identifies the complete document, including its creation time and job ID (jobs with otherwise identical inputs therefore have different IDs).
Comparing jobs
registeel-diff compares a baseline manifest with a candidate manifest. It reports changes to configuration, code, Python packages, and input-file contents. It omits timestamps, job IDs, schema versions, branches, and selected machine-local paths by default. Content-based list matching prevents one added pipeline stage from producing false changes in every later stage.
The governing procedure selects the baseline and determines which differences are acceptable.
| Status | Meaning |
|---|---|
0 | No non-exempt differences |
1 | One or more differences |
2 | A manifest was missing, invalid, or corrupted |
Current integrations
| Repository | Manifest created | Manifest record |
|---|---|---|
| Minccino | Before recipe stages execute | Manifest stored in the job directory; ID added to the job record |
| Machamp SFT | Before the model loads | Manifest and ID stored with the Aim run; training manifest can reference the Minccino manifest |
Other Machamp training pipelines do not currently emit Registeel manifests.