| English | 中文 |
This prompt is for handoff after review; it does not establish an LLM call or generated-code event. Supply the plan, comparison-plan.json, development-profile.json and code-review checklist first.
Build a reusable data-plus-regression-skill composition from the attached scientific design. Current scope permits checking and presenting the plan only. Without actual review for this plan, do not generate experiment callers or execute models, and do not reuse an earlier delegation.
First explain sources, data-specific reasons, alternatives and uncertainty for V01–V11, P01–P07 and H01–H08. Compare A’s eight environmental variables with B’s identical set plus latitude/circular longitude; distinguish fixed-setting paired contrasts from equal-budget tuned-pipeline contrasts.
After real approval, present skill adaptation interfaces and synthetic checks first. Reusable preparation, spatial grouping, development-only fitting and outer refitting belong in skills; workflows retain orchestration and reasons. Copying the existing single-task controller does not implement nested spatial validation.
Before code generation, freeze common sample keys, spatial protocol, four parameter combinations, seeds, metrics and outer protection. Create actual input/task plans under the existing review contract, retaining hashes and real reviews. Never treat the multi-arm draft as an approved single-task plan.
Save actual caller code, normalized tasks, approved plans, prompt, input/fold hashes and per-decision field/code mappings in gen_script with the new external project result. Distinguish copied templates, reused skills and newly agent-authored code; do not invent historical reasoning.
Review unchanged sources, paired keys, circular longitude, forest-sum missingness, fold-specific fitting, effective defaults/seeds and outer isolation. Run synthetic checks/preflight before authorized execution. Stop unsupported parts and explain required adaptation.
Report published split versus transfer, fixed-setting versus independently tuned contrasts, regional errors and seed variability separately. Original test scores have been seen; do not claim new independent validation. Without a regional grid, do not promise map or block-artefact checks. Preserve old results, use a fresh external project directory, and never inspect credentials or put private paths in shared materials.
Generation/execution permissions remain false for this round. A review response must distinguish plan changes from authorization to generate, adapt skills or execute within defined scope. This prompt itself grants no authorization.