[frazane]

Registered experiments live at /store_new/mch/msopr/ml/experiments, mirroring devml's model registration (immutable copies, append-only JSONL index, tombstoned names).

• checkpoint: accepts a registered model name, e.g. amber-ridge
• evalml register|unregister|list|publish manage the store, the model-store cross-links, and the Experiments Confluence page
• the workflow copies the config into the results directory, so results are registerable as-is

Details in docs/experiment-store.md.

Note

This technical implementation gives us some basic elements on which we can develop our processes. It only does the bare minimum we needed (experiments tracking and cross-references with registered models) and it's a temporary solution. In the future we will adopt the anemoi-nexus.

[Louis-Frey]

What is the state of this PR? It would be nice to be able to use it from main.

[frazane]

@copilot resolve the merge conflicts in this pull request

[frazane]

@Louis-Frey with your approval, I think we could merge this.

[Louis-Frey]

I didn't have time to look at the code (and don't have now), but I used the branch twice and it seems to be doing what it's supposed to do.

Good to merge from my side.

[Copilot]

🟡 Changes recommended

Unresolved concurrency races can discard successful registrations, allow duplicate evaluations, and overwrite newer catalogue data, while the inference patch no longer overrides legacy checkpoints’ active input paths and metadata.

6 open issues

• Keep legacy inference config keys for pinned checkpoints · New
• Serialize duplicate registrations across different names · New
• Prevent rebuild from discarding concurrent registration events · New
• Resolve dangling symlink targets relative to the source directory · New
• Archive the explicitly selected config.yaml · New
• Prevent stale inventory snapshots from overwriting newer catalogues · New

What changed in this PR

Adds a shared experiment archive to EvalML, connecting completed evaluations with registered models and publishing an inventory to Confluence.

Changes:

• Adds registration, unregistration, listing, index repair, and publishing commands.
• Supports registered model names as checkpoints and preserves experiment configs alongside results.
• Adds store lifecycle tests, documentation, and updated example configurations.

File	Description
workflow/​tools/​config.schema.json	Documents registered model names as checkpoint inputs.
workflow/​Snakefile	Copies the experiment config into results.
workflow/​rules/​common.smk	Preserves registered model names in run identifiers.
tests/​unit/​test_store.py	Tests store lifecycle, safeguards, and rendering helpers.
src/​evalml/​store.py	Implements archiving, indexing, model links, and Confluence publishing.
src/​evalml/​registration.py	Extracts configuration metadata and registration provenance.
src/​evalml/​config.py	Resolves registered model names to checkpoint paths.
src/​evalml/​cli.py	Exposes experiment-store commands.
resources/​inference/​configs/​sgm-forecaster-global-ich1.yaml	Switches the inference metadata patch to the multidataset variant.
docs/​experiment-store.md	Documents commands, storage guarantees, and publishing.
config/​forecasters-ich1-oper.yaml	Uses the registered amber-ridge model.
config/​forecasters-ich1-oper-fixed.yaml	Uses the registered amber-ridge model.
.gitignore	Ignores .local and .claude paths.

_{Review effort: Balanced}

[lclanzi]

Very useful feature, thanks!