trace

# Trace: Concept Lineage Deep Dive

Trace the complete research genealogy of a single concept. Answers "Where did this come from? How did it evolve?"

## When to Use

- "Trace [concept]"
- "Where did [concept] come from?"
- "Show me [concept]'s paper history"
- "History of Attention mechanism"
- Understanding one concept's full evolution

## When NOT to Use

- Learning multiple concepts → use `deep-dive`
- General domain exploration → use `domain-vocab`
- Latest research only → use `frontier`

## Core Value

> When you dig deep into one concept, you see the entire field.

Trace provides **vertical depth** (one concept, complete history) vs. deep-dive's **horizontal breadth** (many concepts, key papers).

## Workflow

### Phase 1: Concept Identification

**Input:** Concept name (optionally with domain context)

**Actions:**
1. Confirm the concept and its domain
2. Identify potential ambiguity (e.g., "Attention" in NLP vs. psychology)
3. Prime context with domain-vocab tokens (light version)

**Output:**
```yaml
concept: "Attention Mechanism"
domain: "NLP / Deep Learning"
disambiguation: "Neural attention for sequence models, not cognitive attention"
search_keywords: ["attention", "neural machine translation", "sequence to sequence"]
```

### Phase 2: Root Paper Discovery

**Objective:** Find the seminal paper that introduced this concept

**Actions:**
1. Search Semantic Scholar + arXiv for concept
2. Filter by:
   - High citation count
   - Early publication date
   - Title/abstract directly mentions concept introduction
3. Verify with heuristics:
   - Does the abstract say "we propose/introduce"?
   - Is it widely cited as origin?
4. If multiple candidates, ask user to confirm

**Decision Matrix:**

| Signal | Indicates Root | Score |
|--------|----------------|-------|
| "We propose/introduce X" in abstract | Strong | +3 |
| Published before all high-citation papers on topic | Strong | +3 |
| 1000+ citations | Likely influential | +2 |
| Cited by survey papers as o