| English | 中文 |
| Layer | Responsibility | Entry |
|---|---|---|
| Method | Target definition, feature roles, DAG identification, regression and interpretation assumptions | Regression method |
| Task skill | Reviewed target/feature preparation, fitting, evaluation, predictive explanation and supported causal estimation | Regression skill |
| Model families | Linear regression, tree-based models and sequence references | Model directory |
| Common evaluation | Training-mean reference and shared candidate checks | Evaluation |
| Software demonstration | Direct CPU synthetic Transformer input/training/reload cases; not a registered tool | Software cases |
| Framework skill | mlr3 environment, API and extension guidance | mlr3 |
| Sequence reference | External Python/PyTorch TimeXer instructions; no runner adapter | TimeXer |
| Workflow | General data/skill combinations, DAG design step, branches and handoffs | Regression combination |
| Project | A particular question, experiments, comparisons and review | Project 0001 |
The existing R executable ID remains ml_regression_workflow, with an mlr3 backend. The separate Python regression entry is ml_python_regression_workflow; its language path specifies the distinct task, dependencies and review bundle. mlr3_workflow_reference and timexer_workflow_reference remain non-executable guidance. Package APIs are not tool IDs.
python -m runtime.cli list-tools --category model --query mlr3
python -m runtime.cli describe-tool ml_regression_workflow
python -m runtime.cli check-tool ml_regression_workflow
python -m runtime.cli run-tool ml_regression_workflow --params-file config/examples/run-tool/aq-bench-ml-regression.json --dry-run
python -m runtime.cli run-tool ml_regression_workflow --params-file config/examples/run-tool/synthetic-causal.json --dry-run
Use prediction settings to configure target, features, train/validation/test labels, missingness, bounded rpart/ranger candidates, seeds and explanations. The training-only filter records missingness, variation, correlation and VIF diagnostics. Predictive explanations include permutation importance, PDP, ICE and optional empirical-background permutation Shapley estimates. These explain predictions, not causal effects.
Use synthetic causal settings for one continuous-treatment, cross-fitted partially linear slope. It requires a declared DAG, total-effect adjustment set, manual covariates, complete cases, one nuisance specification and scientific assumptions. It supports independent-unit or independent-group folds; arbitrary causal estimands, IV/DID and automatic causal discovery need adaptation. The known-effect synthetic example is implementation evidence. Observational causal analysis requires a study-specific design.
No project ID selects the regression skill’s local, Git-ignored results/<run_id>/. A project ID selects external assets/results/<project_id>/; out_dir and relative output_files may choose a useful internal layout. The adapter preserves existing files, creates internal .airmosaic/<run_id>/ records/temp and writes artifacts.json for local discovery. Do not publish generated results. Other adapters retain their documented layouts.
Model computation needs no login or network. Set AIRMOSAIC_R_LIBRARY explicitly for a reviewed external package directory containing mlr3, mlr3learners, ranger, rpart and jsonlite; otherwise only standard R libraries are used. Personal startup files and implicit personal/site libraries are disabled. Three private roots remain required. MCP uses the same configuration and checks; execution requires the startup grant --execute-tool ml_regression_workflow.
Classification, other learners, automatic tuning, predictive cross-validation, new-data inference and general experiment orchestration require additional adaptation. The published basis distinguishes source-informed design from numerical paper reproduction. Scientific review and approved data reuse remain separate from execution.
Synthetic workflow checks may explicitly select result_scope=workflow, limited by the asset contract to a fixed Git-ignored workflow result tree and manifest-listed synthetic inputs. Scientific projects retain external results; the default scope is unchanged.