07. Decision-Value Experiment Planner
Purpose
设计有决策价值的实验。实验不是为了填表、补齐口径或让 ablation 看起来完整,而是为了回答明确假设或支持决策。
Experiment modes
先判断当前模式:
Exploration mode:
实验用于判断 claim 是否值得继续。
Evidence mode:
实验用于支撑 residual claim。
Reviewer-defense mode:
实验用于回应合理 reviewer 质疑或补投稿完整性。
Decision-value 原则在三种模式都成立,但“决策价值”的定义不同。
Required fields for every experiment
Experiment ID:
Problem addressed:
Which part of the stated problem does this experiment answer?
Hypothesis:
Decision this experiment informs:
Minimum valid setup:
Invalid cheap proxy to avoid:
Expected outcomes:
Decision for each outcome:
Stop condition:
Valid minimal experiment
初期探索应使用少量但有效的 decisive experiments。可以缩小范围,但不能改变问题本质。
合格做法:
- 选择代表性任务,而不是十个任务;
- 只比较关键 baseline,而不是完整表格;
- 在小规模但机制不失真的 setting 下完整训练到合理收敛;
- 保留 claim 所依赖的关键变量;
- 控制最重要 confound;
- 结果足以决定 continue / pivot / kill。
不合格 cheap proxy:
- 本来需要 40 epoch 收敛,只跑 10 epoch;
- 用 MLP / simple attention 替代真正要测试的复杂模块;
- 用 toy setting 替代 claim 所依赖的真实 setting;
- 用弱 baseline 证明自己有效;
- 用小到失真的 dataset 支撑 scaling claim。
Evidence categories
Necessary evidence:
没有它,claim 站不住。
Supporting evidence:
增强说服力,但不是必须。
Cuttable evidence:
看似完整,但信息增益低,会膨胀论文。
Future work:
有价值,但不属于当前 paper unit。
Stop low-information-gain experiments
如果一组实验预计不会改变以下任何决策,就不做:
- 是否继续这个方向;
- 是否修改 main claim;
- 是否改变 paper type;
- 是否补强 reviewer defense;
- 是否指导后续方法改进。
尤其当实验只是重复确认“不可行 / 无效果”,且不会改变决策时,应停止。
User explicitly asks for full comparison
如果用户明确要求补满对比实验,原则不阻止执行,但要标记实验模式为 reviewer-defense / completeness mode。
仍需问:
这个对比是否是合理 reviewer 会要求的?
它是否覆盖真实 competing method?
它是否会改变 paper 可信度?
Output
写入:
06_experiments/decision_value_plan_Cxxx.md
06_experiments/experiment_log.md
Exit condition
可以进入 reviewer risk check 或 execution,当:
- 每个核心实验有 hypothesis、decision、minimum valid setup、stop condition;
- necessary / supporting / cuttable evidence 已区分;
- 没有低信息增益穷举;
- 没有用失真的 cheap proxy 作为最小实验。
Pipeline integration
When used inside the modular pipeline, finish by updating or creating the relevant research-state artifacts. Do not rely on chat context as the only source of truth.
State artifacts to update:
research_state/06_experiments/decision_value_plan_Cxxx.mdresearch_state/06_experiments/experiment_log.md
Recommended next step:
- Continue with
reviewer-risk-check, unless the exit condition is not met.
If information is missing, make the smallest explicit assumption that allows progress, record it in research_state/00_state/current_state.md, and mark the status as provisional.
Problem alignment rule
If an experiment improves a metric but does not clarify the stated problem or support/refute the active claim, mark it as low decision value. Experiments should answer the problem, not merely decorate the paper.