Back to Details

experiment-logbook

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Experiment Logbook

实验日志本

Use this skill to keep experiments scientifically interpretable. The experiment record is part of the research output.

使用此技能可确保实验具备科学可解释性。实验记录是研究产出的一部分。

Read First

必读内容

```
references/experiment-policy.md
```
```
references/output-contracts.md
```

```
references/experiment-policy.md
```
```
references/output-contracts.md
```

Before Running

运行前记录

Record:

hypothesis
editable scope
frozen evaluation harness
metric and direction
data version
time or compute budget
command
expected output
stop condition

记录以下内容：

假设
可编辑范围
固定的评估框架
指标及优化方向
数据版本
时间或计算资源预算
命令
预期输出
停止条件

Run Record

运行记录

Update

experiments/registry.csv

and create

experiments/<run_id>.md

for non-trivial runs.

Minimum fields:

run_id
git commit
command
config
seed
dataset
metric
result
runtime
status
notes

更新

experiments/registry.csv

，并为重要运行创建

experiments/<run_id>.md

文件。

必填字段：

run_id
git commit
命令
配置
seed
数据集
指标
结果
运行时长
状态
备注

Autonomous Loop Rules

自主循环规则

Establish a baseline first.
Keep a machine-readable ledger.
Commit candidate changes before each run when code changes matter.
Keep only improvements if that is the agreed policy; otherwise record all variants.
Never alter the metric, data split, or evaluation harness silently.
Stop for human approval if the change alters scientific meaning.

首先建立基准线。
维护机器可读的分类账。
当代码变更产生影响时，在每次运行前提交候选变更。
若已约定相关规则，仅保留改进结果；否则记录所有变体。
不得擅自修改指标、数据划分或评估框架。
若变更会改变科学含义，需暂停并等待人工审批。

Output Separation

输出分离

```
train_outputs/
```
: trusted training evidence.
```
explore_outputs/
```
: exploratory variants and caveated ideas.
```
analysis_outputs/
```
: result analysis and figures.
```
debug_outputs/
```
: failure diagnosis.

```
train_outputs/
```
: 可信的训练证据。
```
explore_outputs/
```
: 探索性变体及带有说明的想法。
```
analysis_outputs/
```
: 结果分析及图表。
```
debug_outputs/
```
: 故障诊断记录。