先看這個問題為什麼值得解Why this problem is worth solving
情境先行Start with the situation
藥物模型在隨機切分後表現很好，但相近的分子骨架同時出現在訓練集與測試集。模型可能只記住分子家族，未必能推廣到新的化學空間。 A drug model performs well after a random molecular split, but related scaffolds occur in both train and test. It may memorize families rather than generalize to new chemical space.
在科學與醫療場景，資料洩漏或錯誤外推不只影響排行榜，也可能浪費實驗資源或造成不安全決策。 In science and health, leakage or invalid extrapolation can waste experiments or lead to unsafe decisions, not merely distort a leaderboard.
先備知識Prerequisite
- 理解相關性、驗證與資料切分 Understand correlation, validation, and data splitting
讀完你能做到After this path
- 區分預測結果、機制假說與介入證據 Distinguish prediction, mechanistic hypotheses, and intervention evidence
- 辨識實體、時間、場域與家族之間的資料洩漏 Identify entity, temporal, site, and family leakage
- 為高風險用途建立外部驗證與失敗條件 Define external validation and failure conditions for high-stakes use
閱讀框架Concept frame
- 核心問題Question
- 模型在歷史資料上預測得準，何時仍不能支持新的科學或醫療結論？ When does accurate prediction on historical data still fail to support a new scientific or health conclusion?
- 對照軸Comparison axis
- 驗證資料與預期使用場景的距離 Distance between validation data and intended use
- 判讀方式Evidence to inspect
- 並排閱讀三種相對情境，再回到論文檢查消融實驗、反事實測試與適用範圍。Read three relative scenarios side by side, then inspect the paper's ablations, counterfactual tests, and scope statements.
回答後查看解析Answer, then view the explanation
你的預測Make a prediction
要評估模型能否推廣到未見的分子家族，哪個切分最有資訊？ Which split is most informative for generalization to unseen molecular families?
看現象，不急著講故事Observe before narrating
三個診斷訊號Three diagnostic signals
這些是待檢查的診斷假說。下一步用條件比較、消融實驗或反事實測試排除其他解釋。These are diagnostic hypotheses to test. Use condition comparisons, ablations, or counterfactual tests to rule out alternatives.
一次看完三種情境Read three scenarios side by side
三種情境對照Three-scenario comparison
驗證資料離實際用途愈遠時，對預測、發現與實際使用的主張應如何依序收斂？ As distance grows, how should prediction, discovery, and deployment claims be downgraded?
對照軸：Comparison axis: 驗證資料與預期使用場景的距離 Distance between validation data and intended use
可評估內部重現性，但仍不能直接推論因果或臨床效益。 Internal reproducibility can be assessed, but causal or clinical benefit does not follow automatically.
加入按時間、場域或家族的資料切分，並檢查預測校準。 Add temporal, site, or family splits plus calibration analysis.
先把結果視為產生假說的線索，等待外部或前瞻驗證。 Downgrade the result to hypothesis generation pending external or prospective validation.
從直覺到機制From intuition to mechanism
把機制講清楚Explain the mechanism
模型可以是顯微鏡、篩選器或導航，但工具看見的規律不會自動變成自然定律或治療效果。 A model can be a microscope, filter, or navigator, but a pattern seen by a tool does not automatically become a natural law or treatment effect.
模型讀入實驗、觀測或臨床資料及其測量流程，產生預測、候選項目或待驗證假說；用途是在明確的科學與使用條件下提升研究和決策效率。 Input: experimental, observational, or clinical data plus measurement procedures. Output: predictions, candidates, or hypotheses. Objective: improve information and decision efficiency under explicit scientific or use conditions.
- 01 先說明測量方式，以及真正關心的科學目標。 Define measurements and the scientific target of interest.
- 02 依實際外推單位設計資料切分、比較基準與不確定性分析。 Build splits, baselines, and uncertainty around the real extrapolation unit.
- 03 以獨立實驗、外部群體或前瞻流程驗證候選。 Validate candidates through independent experiments, external cohorts, or prospective workflows.
把回溯資料中的關聯，寫成作用機制、因果效果或可直接採用的醫療建議。 Writing a retrospective association as a mechanism, causal effect, or deployable medical recommendation.
閱讀科學 AI 主張時，應記下測量方式、切分單位、外推範圍、驗證層級與失敗成本。 An AI-for-science claim should state its measurement, split unit, extrapolation scope, validation level, and failure cost.
不用回看頁面的自我檢查Check your recall without looking back
三個回想題Three recall questions
先口頭回答，再展開解析。Answer aloud before opening each explanation.
01 為何隨機切分可能過度樂觀？ Why can a random split be optimistic?
同一實體、家族、場域或時間鄰近樣本，可能同時出現在訓練集與測試集。 Samples from the same entity, family, site, or nearby time can appear across train and test.
02 預測與機制有何不同？ How does prediction differ from mechanism?
預測捕捉可用關聯；機制還要說明現象如何產生，並接受介入測試。 Prediction captures useful association; mechanism additionally concerns the process generating a phenomenon and intervention-based tests.
03 外部驗證回答什麼？ What does external validation answer?
方法在不同來源、時間或族群下是否仍具校準與效用。 Whether calibration and utility persist across sources, time periods, or populations.
把工具帶回你的研究Take the tool back to your research
最後一個任務One final task
選一篇科學或醫療 AI 論文，只依書目資料與摘要列出測量方式、切分單位、驗證方式與預期用途；沒有明說的項目寫『未確認』。 Choose an AI-for-science or health paper and, from metadata or abstract only, list measurement, split unit, validation, and intended use; mark unstated items unverified.