--- name: design-diagnostic-analysis description: >- Design diagnostic analyses that explain why a method works, when it fails, and which mechanism supports the paper's claims. Use when main results are present but the causal story, mechanism, or failure modes are unclear. Output is an analysis protocol with metrics and expected interpretations. metadata: owner_model: ParadoxGPT-Designer-4B skill_family: designer version: "0.1" output_type: generation default_language: zh trigger_keywords: [diagnostic analysis, mechanism, why it works, failure mode, analysis experiment, 机制分析, 诊断实验] required_inputs: [method_claim, main_results, uncertain_mechanism] optional_inputs: [method_summary, current_experiments, constraints] handoff_to: [design-baselines-and-ablations, map-claims-to-experiments, check-claim-evidence-alignment] --- # Design Diagnostic Analysis ## Purpose 设计分析实验来解释方法为什么有效、何时失败、机制证据是否支持论文 claim。这个 skill 处理主实验之外的“解释性证据”,避免论文只有 leaderboard 结果却没有可审稿的机制闭合。 ## When to Use - 主结果不错,但 reviewer 可能问“为什么有效”。 - paper 有 mechanism claim,但缺分析实验。 - 方法失败模式不清,claim 边界无法校准。 ## Do Not Use - 需要 baseline / ablation 结构 → 用 `design-baselines-and-ablations`。 - 不知道哪些 claim 需要实验 → 用 `map-claims-to-experiments`。 - 只是检查已有 claim 是否有证据 → 用 `check-claim-evidence-alignment`。 ## Required Inputs - `method_claim`: 需要解释的 claim。 - `main_results`: 主实验结果摘要。 - `uncertain_mechanism`: 当前不确定的机制或失败点。 - `method_summary`: optional。 ## Output Contract 输出必须至少包含: 1. **diagnostic_questions** - 分析实验要回答的问题。 2. **analysis_protocol** - 每个诊断实验的设置、对照、变量。 3. **metrics** - 应观察的指标,不只主任务分数。 4. **expected_interpretation** - 不同结果分别支持/削弱什么 claim。 5. **failure_mode_probe** - 至少一个失败模式或边界测试。 ## Procedure 1. 把 method_claim 改写成可被诊断的问题。 2. 找出不确定机制:信息来源、组件作用、数据分布、搜索过程、反馈循环等。 3. 为每个机制设计可观测 proxy 或对照。 4. 指定 metrics,区分 outcome metric 和 mechanism metric。 5. 写出如果结果相反,应如何调整 claim。 ## Quality Bar 一个好的输出必须: - diagnostic question 直接对应 mechanism claim。 - protocol 可执行,变量控制清楚。 - expected_interpretation 包含正反两种结果。 - failure probe 能帮助校准 claim 边界。 一个差的输出: - 把更多 benchmark 当成机制分析。 - 只说“做 case study”,没有指标和解释规则。 - 结果无论怎样都被解释为支持论文。 ## Failure Modes - **confirmation bias**:只设计会支持自己的分析。 - **proxy 不相关**:指标看似可测但不能解释机制。 - **分析不可复现**:只放 cherry-picked case。 - **边界缺失**:不测试方法失败条件,claim 无法校准。 ## Handoff - 如果还缺 baseline/ablation → `design-baselines-and-ablations` - 如果 claim 到实验的整体闭合不清 → `map-claims-to-experiments` - 如果要审计已有证据是否支撑 claim → `check-claim-evidence-alignment` ## Example Input: ``` method_claim: gains come from iterative correction, not just more tokens main_results: multi-round agent outperforms single-round baseline uncertain_mechanism: round-level feedback vs token budget ``` Output: ``` diagnostic_questions: Q1: 多轮收益是否来自反馈纠错? Q2: 哪类错误最能被后续轮次修复? analysis_protocol: D1: same-token single response vs multi-round with feedback. D2: remove feedback but keep same number of rounds. D3: label error trajectories across rounds. metrics: - correction rate per round - repeated-error rate - success conditional on first-round failure expected_interpretation: - 如果 D2 接近完整方法,反馈机制 claim 变弱。 - 如果 first-round failure recovery 高,支持 iterative correction。 failure_mode_probe: - 测试需要全局规划的任务,观察多轮局部修正是否失效。 ```