| English | 中文 |
Workflows explain reusable combinations of data and skills: purpose, input contracts, ordered handoffs, alternative branches, checks and review. Code and concrete software operations belong in skills; scientific reasoning belongs in methods; specific scientific questions and experiments belong in projects.
See the regression combination, including target definition, variable roles and DAG design. Its referenced registered capability is executable, while the recipe YAML is documentation. There is no general YAML executor, project loader, automatic experiment orchestration or scientific approval service.
The model reproduction combination adds source audit, frozen candidate/reference snapshots, comparable views and fresh project comparison batches. compare_existing uses a reviewed direct helper with default preflight; reproduce_and_compare is agent-guided and requires explicitly selected skills. Its prompt describes handoffs and review branches. Source collection is reviewed separately through an authorized route; no general remote execution or registered comparison adapter is provided.
Generic demonstrations use adapter-supported ignored results, while scientific experiment outputs belong to the corresponding project. Current runners retain their existing external compatibility locations; direct project-directory output support is not implemented. Human review precedes any inclusion in assets/. Preserve historical outputs. See the project guide and result and review contract.
The reviewed regression adapter additionally supports explicit result_scope=workflow (--result-scope workflow in the CLI). Only hash-verified manifest-listed repository synthetic CSVs are accepted. Outputs stay in Git-ignored workflows/ml_regression/results/<timestamp>_<task>/, with receipts/temp under .airmosaic/<run_id>/ in that result root. Project IDs, real/external data, paths outside this fixed tree and nested repositories are rejected. The auto default is unchanged; scientific studies retain external project results. Reviewed workflow orchestration examples are reusable source; per-run generated code and design basis stay in the result’s gen_script/, following the common handoff rules. This check provides neither scientific approval nor OS isolation.