| English | 中文 |
The request instantiates the workflow prompt. The reviewed controller uses the registered regression skill and public synthetic task. It replays agent-authored code; no LLM generation API is called during execution.
For a literature-grounded real-data plan before code generation, see the AQ-Bench paired coordinate design and decision-to-code review map. That case is a pending design with task drafts, not an executed spatial validation; the synthetic replay below retains its own scope.
From the repository root, explicitly prepare the separate storage roots and R library using the skill, then run:
python workflows/ml_regression/examples/generated_regression.py --task-name synthetic_regression
python workflows/ml_regression/examples/generated_regression.py --task-name synthetic_regression --utc-offset +08:00 --execute
Preflight does not fit or create result folders. Execution reserves workflows/ml_regression/results/<YYYYMMDDHHmm>_<task-name>/ exclusively. The name uses minute precision; the explicit offset defaults to +08:00. Internal --run-id remains optional receipt correlation, not the visible directory name. Task names must be public lowercase slugs. Conflicting names and existing internal receipts are rejected.
Each result contains:
gen_script/run.py, task.json, prompt/request copies, paired generation*.md and structured generation.json;gen_script/dependencies.json and dependencies.patch, identifying reused public code and preserving changes/additions against a Git base, rather than copying a full source tree;events.jsonl, separate runner receipts, model/prediction/explanation artifacts and independent verification.json;retrospective*.md, overall terminal workflow.json and delivery-sha256.json.No new source_snapshot/ or input-data copy is produced. The ignored result tree retains all local files; historical runs are not automatically renamed or removed; explicitly authorized redundant synthetic batches may be cleaned after review. The current checked receipt summarizes the retained batch; an authorized cleanup manifest records removed redundant synthetic batches.
The generation basis is explicit design information saved before computation, not a recreated historical thought trace. Events capture actual controller boundaries; preflight is buffered until a directory is reserved. Full host-agent conversations, generated-code edit history and inner R operations are not automatically captured. Retrospectives are labelled as written afterward. Follow the common record contract.
To replay an executed program, use its saved relative path, replacing the example directory below:
python workflows/ml_regression/results/<timestamp>_<task>/gen_script/run.py --task-name synthetic_replay --execute
Replay uses the saved task and generation materials, checks current dependency hashes and creates a fresh result directory. Changed dependencies require review and reconstruction/adaptation; patches are never applied automatically. They preserve LF-normalized public text, with separate byte and normalized hashes, including empty additions and missing terminal newlines.
Inspect overall workflow.json and the delivery hashes as well as the separate process receipt. A successful R process alone does not establish complete workflow delivery. The fixed smoke criteria include exact validation/test keys and truths, independently recomputed selected and training-mean baseline metrics, reload equality and test RMSE below the synthetic baseline. Negative engineering checks cover paths, private/external inputs, records, dependency reconstruction and failed verification/sealing. These checks do not approve scientific claims or establish performance on new real data.