| name | research-experiment |
| description | [Read when prompt contains /research-experiment] |
| metadata | {"openclaw":{"emoji":"🧪","requires":{"bins":["python3","uv"]}}} |
Don't ask permission. Just do it.
| File | Source |
|---|---|
project/ | /research-implement |
plan_res.md | /research-plan |
iterations/judge_v*.md | /research-review（最后一份 verdict 必须是 PASS） |
验证 PASS： 读取最新的 judge_v*.md，确认 verdict: PASS。如果不是，STOP。
| File | Content |
|---|---|
experiment_res.md | Full experiment report (full training, ablations, supplementary experiments) with explicit headline metrics, baselines, guardrails, and figure anchors |
experiment_analysis/analysis_{N}.md | 每轮实验分析报告（迭代过程中产生） |
修改 epoch 数为 plan_res.md 中指定的正式值。不要改代码逻辑，只改 epoch。
cd project && source .venv/bin/activate
python3 run.py # full epochs
记录完整训练的 [RESULT] 输出。
读取训练输出，评估：
根据 plan_res.md 中的消融计划，执行 2-3 个消融实验：
对每个消融：
# Example: 去掉 attention module
python3 run.py --epochs 2 --ablation no_attention
⚠️ 这是 Novix Exp Analyzer 机制 — 对已有结果进行分析，提出补充实验，执行后再分析。
循环 2 次：
读取当前所有实验结果（full training + 消融），写入分析报告 experiment_analysis/analysis_{N}.md：
# Experiment Analysis Round {N}
## 当前结果摘要
- Full training: {metrics}
- 消融实验: {key findings}
## 发现的问题或机会
1. {observation} → 建议: {experiment}
2. ...
## 补充实验计划
| 实验名称 | 目的 | 修改内容 | 预期结果 |
|----------|------|----------|----------|
| {exp_name} | {why} | {what to change} | {expected} |
补充实验类型参考（Novix Exp Analyzer 的典型输出）：
根据分析报告中的计划，修改代码并执行补充实验。只改实验相关参数/配置，不改核心算法逻辑。
cd project && source .venv/bin/activate
python3 run.py --experiment {exp_name}
记录结果后，回到 4.1 进行下一轮分析（共 2 轮）。
汇总所有实验结果（full training + 消融 + 2 轮补充实验），写入 experiment_res.md：
# Experiment Report
## Full Training Results (from execution log)
- Epochs: {N}
- [RESULT] train_loss={value}
- [RESULT] val_metric={value}
- [RESULT] elapsed={value}
- [RESULT] device={device}
- [METRIC] name={headline_metric} value={value} unit={unit} baseline={baseline}
- [GUARD] name={guard_name} value={value} threshold={threshold} pass={true/false}
- [FIGURE] file={figure path}
> 以上数值来自真实执行输出。
## Training Analysis
- 收敛情况: {converged / still improving / diverged}
- 过拟合: {yes/no, evidence}
## Ablation Studies
| 实验 | 修改 | val_metric | vs Full |
|------|------|-----------|---------|
| Full model | — | {value} | baseline |
| No {component} | 去掉 {X} | {value} | {-/+}% |
| ... | ... | ... | ... |
## Supplementary Experiments
### Sensitivity Analysis
| 超参数 | 值 | val_metric | 备注 |
|--------|-----|-----------|------|
| ... | ... | ... | ... |
### Comparison with Baselines
| 方法 | val_metric | 备注 |
|------|-----------|------|
| Ours | {value} | — |
| {Baseline} | {value} | ... |
## Scope / Evidence Boundary
- baseline: {which baseline is used}
- protocol / guardrail: {evaluation rule}
- evidence_type: {simulator / local_runtime / full_runtime}
### Visualizations
- 训练曲线: `project/figures/training_curve.png`
- {其他可视化}: `project/figures/{name}.png`
## Conclusions
- {key findings from all experiments}
## Limitations
- {limitations and future work}
Use this when the user needs to choose between multiple ML routes after survey but before committing to implementation. Compares candidate approaches, selects one, records rejected routes, and keeps a fallback.
Use this when the project needs real baseline results before or alongside the main model. Runs classical or literature-aligned baselines under the same protocol and writes a reproducible baseline summary.
Use this when the project needs a dedicated data-quality review before model review. Checks data reality, split correctness, label health, leakage risk, shape consistency, and mock-data disclosure.
Use this when the user wants a draft paper, figure bundle, README, release page, or experiment artifact reviewed before sharing. Checks evidence binding, claim scope, captions, layout clarity, and release readiness.
Use this when the user wants to improve chart quality, standardize plotting style, regenerate release figures, or add captions/protocol notes. Normalizes fonts, colors, legends, units, and scope notes across Scientify figures.
Use this when the user wants to improve README, docs pages, or microsites so a new reader can understand what the project is, how to use it, what artifacts exist, and what the scope boundaries are within one screen.