| English | 中文 |
Read the study plan first. This page specifies what to inspect after approval; no run.py, model or execution events exist for this round. Reasons are reviewable decision summaries, not hidden reasoning records.
The existing caller writes plan["task"] to gen_script/task.json and copies reviewed orchestration after approval. Execution calls ml_regression_workflow through the skill gate and runner. It does not scientifically select features/parameters from literature or generate new numerical algorithms. Preserve the truthful copy mode in generation.json.
The preparation module reads configured columns, missingness and existing labels; variable selection diagnoses/filters under the declared policy; fitting trains declared candidates and selects by validation error. They do not read papers to establish scientific validity.
A value appearing in code therefore does not establish its justification. Connect the approved plan, actual configuration and effective execution values to check faithful implementation.
| Decision | Configuration or data | Check after generation |
|---|---|---|
| V01–V08 | First eight features in A/B |
Identical environmental set/order; no silent importance/correlation-driven changes |
| V09–V11 | Only B adds lat, lon_sin, lon_cos |
A retains geographic metadata outside predictors; B’s circular transformation is correct |
| P01–P04 | Prepared data, id and derived columns |
Unchanged source hash; independently recomputable five-field sum/angular transform; identical paired keys |
| P05–P07 | missing, feature_selection and fold fitting |
No clipped/imputed outcomes; fit-only medians; no default correlation filter |
| H01–H02 | models[*].parameters |
Exact four mtry/node combinations; no extra search opportunities for B |
| H03–H05 | Tree count, sampling and omitted parameters | Explicit 500 trees, 0.632 without replacement; record unrestricted depth and variance splitting defaults as well as explicit fields |
| H06 | seeds, order and fit records |
Identical effective seeds for paired settings; fold identity, base and derived seeds retained |
| H07 | Inner comparisons and outer scoring | No outer outcomes/scores in selection; recompute pooled metrics from station errors |
| H08 | explanation |
Traceable common background sampling/budget; no geographic PDP that varies longitude sine/cosine independently |
| Spatial protocol | Fold manifests, distances and refits | Never split cells; buffer fitting side; refit all eligible training records after selection rather than retaining an inner model |
Retain decision_id, source_id, study_reason, alternative, uncertainty, config_field, code_location, effective_value and check_evidence per item. These are requirements for actual post-generation evidence, not records already captured here.
comparison-plan.json is a multi-arm scientific draft, not a model-plan.schema.json single-task record. Do not pass it to the gate as an executable plan. The two task drafts only illustrate published-split fields and have no prepared input.
After approval, design reusable skill interfaces for deterministic preparation, paired cohorts, spatial grouping/buffers, development-only fitting and frozen outer refit/scoring. The current runner requires training/validation/test labels and automatically evaluates the selected model. Do not supply outer-test labels to inner calls and describe repeated scoring as isolation.
Subsequent workflow callers compose these skill interfaces and save task/decision mappings; numerical work stays in skills. Validate preparation, folds and interfaces with synthetic inputs, then bind real prepared-data hashes, exact tasks and actual reviews into each skill plan. Execute only after exact-code checks.
The resulting data flow must be inspectable:
comparison-plan.json → actual approved plan → bound prepared input and fold manifests → gen_script/task.json + run.py → skill preparation/fitting/evaluation → independent checks
This is a proposed flow, not a new executable command or completed run.
Without a regional feature grid, discuss station prediction and dependence diagnostics only; scores cannot establish absence of block-like map artefacts.