Migrating to algorithm terminology and typed profiles

Version 4 changes command names, artifact types, descriptor locations and repository dispatch. Do not reuse version 2 or 3 artifacts by editing their manifest version: validate or rebuild their payloads against the new contract. The template emits and reads version 4 envelopes.

From version 3

This is an intentional naming change, with no old-name aliases:

Previous name	Current name
meds-model-core, meds_model_core	meds-algorithm-core, meds_algorithm_core
ModelConfig, MODEL_PROFILES	AlgorithmConfig, ALGORITHM_PROFILES
ModelRepository (mapping-based adapter)	LegacyAlgorithmRepository
SupervisedModelArtifact (and pretrained/finetuned/probe equivalents)	SupervisedAlgorithmArtifact (and corresponding equivalents)
Request field model	algorithm
meds-model, meds-model-* helper commands	meds-algorithm, meds-algorithm-*
input_*_model_dir, output_*_model_dir, *_model artifact types	input_*_algorithm_dir, output_*_algorithm_dir, *_algorithm
Copier model_name, model_slug, model_description	algorithm_name, algorithm_slug, algorithm_description
commands.py with ModelAlgorithm and generic Config	<slug>.py with <AlgorithmClass> and <AlgorithmClass>Config
models/<profile>/model.yaml	algorithms/<profile>/model.yaml

Set the new algorithm_class answer explicitly for acronyms, such as RETAIN. Update imports, scripts, SLURM environment variables (MEDS_ALGORITHM_*), lockfiles and persisted invocation descriptions together. Serialization records concrete class identifiers; old JSON envelopes require an explicit migration. MEDS-DEV still owns model.yaml, its models/ directory, meds-dev-model, model=, and {model_initialization_dir}. Those upstream names are intentionally unchanged.

From older single-profile repositories

1. Replace the Copier profile answer with profiles: [<name>, ...]. Rename packaged to predict_only. Keep the original source/revision metadata in .copier-answers.yml and review the Copier update diff.
2. Implement the selected profile ABCs and associate profiles with concrete classes in IMPLEMENTATIONS. Shared methods appear once where their semantics match.
3. Rename commands according to the table below. Split previously conditional implementations into distinct profile methods, sharing cohesive components where their semantics really match.
4. Update every invocation, manifest producer, output argument, test adapter and SLURM stage selection.
5. Move the MEDS-DEV descriptor to algorithms/<profile>/model.yaml; add one for each declared profile. Register them independently using --profile; the registered name is <name>__<profile>.
6. Supply explicit split lists for inference and prediction. Supervised/fine-tuned prediction receives input_data_dir; probe prediction receives input_embeddings_dir. Scoring and inference use external_index_dir, not the training-label argument.
7. Replace the global IS_STUB boolean with STUB_PROFILES. Run the acceptance suite for every profile. Record the profile in write_run_result(..., profile=...).

Profile	Old command	New command
All	preprocess_data, pretrain where present	Unchanged
supervised	supervised_train, predict	train_supervised, predict_supervised
finetune	supervised_train, predict	train_finetune, predict_finetune
probe	infer, supervised_train, predict	infer_embeddings, train_probe, predict_probe
zero_shot_direct	predict	predict_zero_shot_direct
zero_shot_materialized	infer, predict	infer_trajectories, predict_zero_shot_materialized
predict_only (formerly packaged)	predict	predict_only

inference splits into embeddings and trajectories, with matching input_embeddings_dir/output_embeddings_dir and input_trajectories_dir/output_trajectories_dir. Task-algorithm artifacts split into supervised_algorithm, finetuned_algorithm, and probe_algorithm, with matching input/output argument names. Prediction output remains output_predictions_dir.

Use Copier 9.17.1 or newer; the template uses its yield extension to generate a directory for each selected profile. A single-profile repository may omit CLI --profile; a repository with multiple profiles must select one. There is no ambiguous compatibility alias for the old generic command names.

This revision supplies the interfaces and tests. It does not automatically migrate downstream algorithm implementations or publish a shared runtime package. See the package assessment.

The core package now owns the single normative CONTRACT.md. Generated SPEC.md is a pointer to the installed package version, readable with meds-algorithm-core-spec. The template and generated projects install the alpha from an immutable Git commit in the separate core repository; no PyPI release is required.

Replace imports from meds_model_contract with meds_algorithm_core and remove the generated compatibility package. Core supplies the registration/input-resolution CLI entry points; keep only the algorithm's own CLI in its project configuration. Import shared fixtures and checks from meds_algorithm_core.conformance. For the typed API, pass algorithm provenance explicitly to AlgorithmRepository(..., repo=algorithm_root). Return E2EResult from core's conformance.types, recording all external inputs; the lifecycle checker now reruns every step with forbidden outcomes absent and fresh output paths. Existing algorithm repositories are not migrated by this template change.

From mapping-based commands to typed algorithms

Core 0.4.0a1 introduces the typed API and version 4 names. Mapping-based dispatch remains available as Command/LegacyAlgorithmRepository after updating imports and arguments; it does not automatically gain the new runner behavior. Migrate deliberately:

1. Move command bodies to the selected Algorithm profile methods. Use their typed request fields.
2. Put algorithm settings/defaults in a frozen AlgorithmConfig subclass; test validation and JSON round trips.
3. Replace internal publication contexts with writes to request.output.directory and a returned PayloadMetadata. Remove algorithm-written manifests and truth attachment; the runner owns those.
4. Extend validate_inputs for payload compatibility. Document any override of its default same-algorithm, same-config rule, especially for external pretrained artifacts or legacy manifests without config.
5. Route the CLI and end-to-end adapters through AlgorithmRepository/StepRunner. Verify saved requests/configuration and fitted artifacts work in a fresh process.

See architecture and UML. No private algorithm migration is included here.