/research

Research Skill

Purpose

Read the materialized research source and extract actionable information needed to implement the proposed method. Record the findings as a structured JSON entry.

When to Use

Phase 2 — after the current implementation has been analyzed.

Input

| Parameter | Type | Description | |-----------|------|-------------| | experiment_path | path | `experiments/{research_name}/` |

The research source was materialized into `{experiment_path}` during Phase 0. Its local path is recorded in `{experiment_path}/log.json` under `metadata.research_source`, and a short descriptive label Phase 0 chose is under `metadata.research_source_kind`. The label is free-form (common values: `pdf`, `file`, `git`, `kaggle_notebook`, `kaggle_dataset`, `arxiv`, `huggingface_model`, `html`, `idea`, `other`), but treat it as a hint only — always follow the actual path in `metadata.research_source`.

Inspect that path and read whatever is there:

• A single file (PDF, Markdown, HTML, `.ipynb`, text, …) → read it directly.
• A directory → read the obvious entry points first (`README*`, `*.ipynb`, top-level notebooks or code, `docs/`, dataset descriptions), then skim the rest as needed.
• A text **idea** (`research_source_kind == "idea"`, typically a short `research_source.md`) → read the user's description carefully and turn it into a concrete method plan. Pick a specific algorithm / library that matches the description, define the hyperparameters you will use, and document your interpretation explicitly in the Phase 2 log entry. If the idea is ambiguous, commit to a reasonable default and note the trade-off — do not invent a citation or claim the idea came from a paper.

Do not try to re-fetch the source. If the content is insufficient, note what is missing in the Phase 2 log entry and proceed with the best analysis you can.

Actions

1. **Read the materialized research source** at `metadata.research_source` (falling back to `research.pdf` for legacy experiments) and extract:

• **Method Summary:** 2-3 short paragraphs describing what the paper proposes, what problem it solves, and how it differs from traditional approaches.
• **Pros:** each advantage the paper claims or demonstrates.
• **Cons:** stated or inferred limitations, assumptions, or weaknesses.
• **Implementation Requirements:**
• Required libraries/packages (with versions if specified)
• Required data format or preprocessing
• Required compute resources (GPU, memory, etc.)
• Key hyperparameters to set
• **Compatibility Analysis:**
• Can the method use the same data as the current baseline?
• Does it need different preprocessing?
• Does it output comparable predictions (same format)?
• Can the same metrics be used for comparison?

2. **Append a Phase 2 entry to `{experiment_path}/log.json`** under `phases`:

   {
     "name": "Phase 2: Research",
     "completed_at": "2026-04-17T10:30:00Z",
     "paper": {
       "title":   "CatBoost: Unbiased Boosting with Categorical Features",
       "authors": ["Prokhorenkova et al."],
       "method_summary": "CatBoost is a gradient-boosting framework that handles categorical features natively via ordered target statistics and uses oblivious decision trees to reduce overfitting."
     },
     "pros": [
       "Native categorical handling — no manual encoding needed",
       "Reduces target leakage with ordered boosting",
       "Strong out-of-the-box performance"
     ],
     "cons": [
       "Training slower than XGBoost for small data",
       "More memory intensive"
     ],
     "requirements": {
       "new_dependencies": ["catboost>=1.2"],
       "data_format": "pandas.DataFrame with categorical columns marked",
       "compute": "CPU is sufficient; GPU optional"
     },
     "compatibility": {
       "same_data":    true,
       "same_metrics": true,
       "preprocessing_notes": "CatBoost takes raw categorical columns; do NOT pre-encode them for the new notebook."
     }
   }

Do not overwrite earlier entries; append to the `phases` array.

Output

• `{experiment_path}/log.json` — updated with Phase 2 research entry
• No other files created or modified