research-loop-init

Research Loop Init

Initialize a .research/ directory from an idea file. Run this only once at the start of a research thread.

Preconditions

• Current directory is a git repository.
• The idea file exists and is readable.
• .research/ does not already exist. If it exists, stop and ask whether to resume instead.

Workflow

1. Read the idea file.
2. Clarify the research frame with the user:
   • motivation: why this research matters
   • core hypotheses: what mechanism or intervention may explain the outcome
   • success criteria: what result would support the decision
3. Propose an initial hypothesis tree with 2-4 top-level hypotheses and optional children.
4. Show the draft tree and wait for user confirmation.
5. Check the current git branch. If it is main, master, develop, or another non-experiment branch, suggest creating a branch but do not create one unless the user explicitly provides a branch name.
6. Create:

   .research/
   ├── IDEA.md
   ├── tree.md
   ├── DASHBOARD.md
   └── experiments/
   

7. Write IDEA.md:

   # Research IDEA
   
   ## 动机
   
   [1-2 paragraphs]
   
   ## 核心问题
   
   - [question]
   
   ## 成功判据
   
   [criteria]
   
   ## 初始假设
   
   - H1: [summary]
   
   ## 参考资料
   
   [links or notes from idea file]
   

8. Write tree.md using the hypothesis-tree format:

   # Hypothesis Tree
   
   ## H1: [full hypothesis]
   Status: 待验
   Evidence: (empty)
   Children: H1.1
   
   ### H1.1: [child hypothesis]
   Status: 待验
   Evidence: (empty)
   Parent: H1
   

9. Write DASHBOARD.md:

   # Research Dashboard
   
   **IDEA**: [one-line idea]
   **Active**: [count] hypotheses | **Last**: [YYYY-MM-DD]
   
   ## Active Hypotheses
   
   - H1: [description] (待验)
   
   ## Next Steps
   
   1. [next discriminating experiment or analysis]
   

10. Report the created files and suggest resuming or recording the first experiment.

Invariants

• Use only 待验, 进行中, 被支持, 被推翻 for hypothesis status.
• Status: in tree.md is plain text, never bold.
• Evidence is append-only and starts as (empty).
• Hypothesis IDs are never reused.
• Do not invent weak ablations or low-value experiments just to fill a table.