受控评测与证据呈现原型Controlled evaluation and evidence-view prototype· 2026
MedMirrorMedMirror
把模型差异整理成可复核的实验记录Turning model differences into reviewable experiment records
在限定范围内保留模型原文、引文、失败和待复核项，探索怎样谨慎地解释实验。Retain original answers, quotations, failures, and open review items within a controlled experiment.
案例摘要Case summary
目标用户Audience
希望核对模型输出差异、运行条件与结论依据的研究和 AI 应用构建者。Researchers and AI builders who want to inspect output differences, run conditions, and the evidence behind a claim.
我的角色Role
围绕评测问题、范围与证据标准，与协作者和 AI 工具推进实现、检查与复盘。Worked with collaborators and AI tools on the evaluation question, scope, evidence standards, implementation, checks, and reflection.
交付结果Outcome
历史记录保留受控运行、原文与证据视图；人工定标及专业复核尚未完成，现有观察不构成临床结论或医疗建议。Historical records retain controlled runs, original answers, and an evidence view. Human calibration and professional review remain incomplete; the observations are neither clinical conclusions nor medical advice.
产品洞察Product insight
不同的模型回答只是观察的起点。实验条件、截断与失败、引文归属以及人的复核，都决定了结论究竟可以说到哪一步。Different answers are the starting point of an observation. Experimental conditions, truncation, failures, quotation provenance, and human review determine how far an interpretation can go.
项目简介Overview
围绕合成病例的模型回答设计受控评测，保留运行记录、原始回答、逐字引文与单页证据视图。当前展示评测方法与工程边界，人工定标和专业复核尚未完成，不输出医学偏见结论。A controlled evaluation using a synthetic case, with run records, original model answers, exact quotations, and a single-page evidence view. It presents methods and engineering boundaries; human calibration and professional review remain incomplete, so it makes no medical-bias claim.
技术栈Tech stack
结果证据Evidence
2026-10-26 只读核对 README、历史运行与项目状态说明；本轮未重跑模型、测试或医学评估。README, historical records, and status documentation were reviewed on October 26, 2026. No model calls, tests, or medical evaluations were rerun for this update.
历史实验使用合成病例，保留模型回答、截断与追问记录，不能推断真实临床效果。Historical experiments use a synthetic case and retain answers, truncation, and follow-up records; they do not establish real-world clinical effects.
人工定标与专业复核尚未完成；工程测试和原文可定位不能替代语义与医学判断。Human calibration and professional review remain incomplete. Engineering checks and locatable quotations cannot replace semantic or medical judgment.
关键决策Key decisions
把截断、失败和完整回答混在一起比较，会让结果失去清晰的解释范围。Comparing failures, truncations, and complete answers together obscures what the experiment can explain.
保留每次运行的条件和状态，将失败与截断单列，再回到原文核对观察。Retain conditions and status for every run, separate failures and truncations, and inspect the original text.
可定位的引文可能被过度解释成已经验证的研究结论。A locatable quotation can be overinterpreted as a verified research finding.
区分原文、机器提取和人工判断，尚未定标的内容保持待复核。Separate original text, machine extraction, and human judgment; keep uncalibrated items pending review.
想看完整的代码和实现细节？Want the full code and implementation details?
在 GitHub 上查看View on GitHub