What are you trying to predict or optimize?
- A) Classification (binary or multi-class)
- B) Regression (continuous value)
- C) Ranking
- D) Other: [specify]

What metric defines success? What target value?

What data is available?
Are there any known data quality issues?
Any leakage risks I should watch for?

Any constraints I should know about?
- Compute budget (GPU/CPU, time limits)
- Interpretability requirements
- Latency requirements for inference
- Regulatory/compliance needs

What validation strategy should I use?
- A) Random split (if data is i.i.d.)
- B) Stratified split (for imbalanced classes)
- C) Time-based split (for temporal data)
- D) Group-based split (to avoid leakage)
- E) Use provided test set (e.g., Kaggle)

What's explicitly out of scope?
Any approaches you want me to avoid?

## Proposed PRD

**Project:** [name]
**Problem:** [what we're solving]
**Goal:** [high-level goal]

**Success Criteria:**

- [ ] Metric > threshold
- [ ] Constraint satisfied

**Constraints:**

- [constraint 1]
- [constraint 2]

**Evaluation:**

- Metric: [metric name]
- Validation: [strategy]

**In Scope:**

- [item 1]
- [item 2]

**Out of Scope:**

- [item 1]
- [item 2]

---

Does this look right? Any changes needed?

When you're ready, say "/start" to begin.

{
  "project": "...",
  "description": "...",
  "created_at": "[timestamp]",
  "status": "approved",
  "problem": "...",
  "goal": "...",
  "success_criteria": ["..."],
  "constraints": ["..."],
  "evaluation": {
    "metric": "...",
    "validation_strategy": "..."
  },
  "scope": {
    "in_scope": ["..."],
    "out_of_scope": ["..."]
  }
}

{
  "status": "running",
  "current": {
    "phase": "ORIENT",
    "iteration": 0,
    "hypothesis_id": null,
    "experiment_id": null,
    "started_at": "[timestamp]"
  },
  "stats": {
    "iterations": 0,
    "hypotheses_tested": 0,
    "hypotheses_validated": 0,
    "hypotheses_rejected": 0,
    "best_score": null
  }
}

{
  "hypotheses": []
}

ORIENT → RESEARCH → HYPOTHESIZE → EXECUTE → ANALYZE → VALIDATE → DECIDE

User: I want to tackle the Kaggle fraud detection competition

Ralph: Great! Let me understand the problem.

What's the evaluation metric for this competition?
- A) AUC-ROC
- B) F1 Score
- C) Log Loss
- D) Other

User: AUC-ROC

Ralph: Got it. Are there any constraints I should know about?
- Interpretability requirements?
- Compute limitations?
- Approaches to avoid?

User: Must be interpretable for compliance. No neural networks.

Ralph: Understood. Here's the PRD I propose:

## Proposed PRD

**Project:** kaggle-fraud-detection
**Problem:** Binary classification of fraudulent transactions
**Goal:** Achieve top 5% on Kaggle leaderboard

**Success Criteria:**
- [ ] AUC-ROC > 0.95 on public leaderboard
- [ ] Model is interpretable (SHAP explanations available)

**Constraints:**
- Must be interpretable (no black-box models)
- No neural networks

**Evaluation:**
- Metric: AUC-ROC
- Validation: 5-fold stratified CV (Kaggle provides test set)

**In Scope:**
- Feature engineering
- Tree-based models (XGBoost, LightGBM, CatBoost)
- Ensemble methods
- SHAP explanations

**Out of Scope:**
- Neural networks
- External data

---
Does this look right? When ready, say "/start".

User: Looks good. /start

Ralph: PRD approved. Starting execution.

[Creates prd.json, ralph.json, backlog.json, log.jsonl, inbox.json]

Beginning ORIENT phase...