Contract API

The normative software interfaces define the public classes, arguments, label-access rules and artifact compatibility requirements. Import them directly from meds_algorithm_core. There is no generated compatibility package.

Module	Responsibility
algorithms	Algorithm/profile ABCs and ordered lifecycles
requests	Typed per-step inputs and CLI field mappings
configuration, serialization	Algorithm-owned config base, strict versioned JSON round trips
runner	Shared execution, validation, staging, measurements and publication
contracts.commands	Derived compatibility views and legacy dispatch
artifacts	Manifests, publication with rollback, upstream references and run results
runtime	Explicit algorithm provenance, memory measurement and subprocess execution
data	Strict split selection, key-only index reads and permitted context targets
tasks.spec	Task identity and relaxation ledgers
evaluation.predictions	Keyed probabilities and post-score truth attachment
integrations.meds_dev	Per-profile registration
conformance	Reusable assertions, synthetic fixtures and explicitly selected integration tests

See the architecture and UML for profile inheritance and request/artifact types. Pass the algorithm checkout as repo= to AlgorithmRepository or StepRunner; unavailable provenance is recorded as null. Algorithm methods return PayloadMetadata, while the runner returns a typed published artifact. Low-level measured_artifact remains for legacy callers; new methods must not publish twice. Use explicit splits for outcome attachment. Install the data extra for dataframe helpers and the conformance extra for shared tests. The base package does not import data or algorithm frameworks.

meds-algorithm-core-spec prints the installed specification. meds-algorithm-resolve-input <artifact> <input_argument> resolves a recorded upstream path and checks its envelope type; the consumer still validates scientific compatibility. Core also supplies meds-algorithm-add-to-meds-dev.

Algorithm-owned adapters remain run_e2e(tmp_path, *, profile, **inputs), fitted_codes(artifacts), and prepare_meds_dev_run(meds_dev, tmp_path, *, profile). The shared lifecycle check reruns the selected profile with evaluated labels removed and fresh outputs; adapters must honor those input overrides, record and apply the fixture seed, and produce reproducible keyed scores. Use read_predictions from conformance.assertions to validate all shards. For separate context/evaluation directories, supply evaluation_index_dir to read_context_labels. conformance.reports.assert_implementation_report validates shared evidence and completed profiles independently of unfinished profiles.