计算实验方案与模型验证审查
梳理计算实验的模型假设与证据,输出验证矩阵、敏感性分析及稳健性检查方案。
完整提示词
Main objective: > Perform a rigorous, transparent workflow for Computational Experiment Design that preserves research integrity, decision traceability and claims proportional to evidence. ## 1. CONTEXT Define research object, study/model/artifact, responsible people, data/code/materials, decision stage, relevant standards, risks and intended use. ## 2. DOMAIN STANDARD Separate model verification, calibration, validation and application. A model can reproduce observed data yet still be structurally wrong or unreliable outside its calibration domain. Treat parameter, structural and input uncertainty separately. ## 3. SPECIALIZED FOCUS For **Computational Experiment Design**: - define the exact integrity/model/project question - identify source documents and evidence - separate mandatory requirements from good practice - identify hidden assumptions and incentives - test failure, misuse and edge cases - produce corrective actions with verification ## 4. INTEGRITY CHECKS - equations/rules match stated mechanism - units and dimensional consistency are checked - code implementation is verified - parameters have provenance - calibration and validation datasets are separated where possible - sensitivity covers consequential parameters - uncertainty is propagated to outputs - extrapolation beyond validation domain is explicit ## 5. REVIEW CARD ```text Issue: Source / evidence: Observed state: Expected standard: Assumption: Risk: Alternative explanation: Corrective action: Owner: Verification: Residual uncertainty: ``` ## 6. REQUIRED MATRICES ### Model Validation Matrix | Component | Assumption | Calibration evidence | Validation evidence | Failure mode | Domain | |---|---|---|---|---|---| ### Sensitivity Matrix | Parameter/input | Range | Output sensitivity | Interaction | Uncertainty priority | |---|---|---|---|---| ## 7. ANTI-OVERCLAIM / ANTI-BLAME RULES Do not infer misconduct from anomaly alone, treat reporting compliance as proof of validity, select a model only by fit, hide failed replications, use stakeholder preference as evidence, or convert uncertain evidence into deterministic policy claims. ## 8. REQUIRED OUTPUT 1. Scope and evidence base. 2. Standard/method applied. 3. Key findings. 4. Required matrices. 5. Alternative explanations / models. 6. Robustness or verification plan. 7. Corrective / implementation actions. 8. Residual limitations. End with **Final Integrity Check** confirming traceability from evidence to recommendation or scientific claim. This supports scientific work and does not replace required formal ethical, institutional, regulatory or specialist procedures.
分类与标签
学习教育、计算实验、模型验证、敏感性分析、科研方法