deep-dive

name	deep-dive
description	Integrated domain exploration combining vocabulary learning with research paper genealogy. Use when user says "deep dive [domain]", "explore [domain] deeply", "domain exploration", or wants to learn concepts with their research origins.
version	0.1.0

Deep Dive: Concept + Paper Integration

Learn domain concepts while tracing their research origins. Combines domain-vocab (conceptual vocabulary) with paper-flow (citation genealogy) for contextualized understanding.

When to Use

• "Deep dive into [domain]"
• "Explore [domain] deeply"
• "Domain exploration"
• "I want to learn [domain] concepts and their papers together"
• Learning a new field with academic depth

When NOT to Use

• Quick concept lookup → use domain-vocab
• Single concept history → use trace
• Latest research only → use frontier
• Citation tree only → use paper-flow

Core Value

Learning concepts alone tells you "what". Learning with paper genealogy tells you "why, where, and how it evolved".

Workflow

Phase 1: Domain Entry (domain-vocab)

Objective: Extract core concepts with authentic expert tokens

Actions:

1. Invoke domain-vocab Phase 0-2:

   • Token priming (key figures, literature sampling)
   • Domain identification (scope, depth level)
   • Core concept extraction (20-30 terms)

2. Identify anchor concepts (5-8):

   • High-impact terms that shaped the field
   • Concepts with clear research origins
   • Terms that connect to seminal papers

Output:

domain: "[identified domain]"
depth_level: L2  # Start at practitioner level
concepts:
  - name: "[concept]"
    difficulty: entry|intermediate|advanced
    is_anchor: true|false  # Anchors get paper-flow treatment
    potential_papers: ["keyword hints for paper search"]

User Checkpoint:

"I'll explore the research lineage for these anchor concepts: [anchors]. Proceed?"

Phase 2: Research Genealogy (paper-flow)

Objective: Build citation trees for anchor concepts

Actions:

1. For each anchor concept:

   • Search papers using concept name + domain context
   • Identify root paper (seminal work that introduced/defined the concept)
   • Build citation tree (parents + children)
   • Extract key papers (high-citation, influential)

2. Cross-link papers:

   • Find papers that connect multiple anchor concepts
   • Identify "bridge papers" that link sub-domains

Output per anchor:

anchor: "[concept name]"
root_paper:
  title: "[paper title]"
  authors: "[authors]"
  year: YYYY
  key_contribution: "[one sentence]"
citation_tree:
  parents: [papers this cites]
  children: [papers citing this]
  key_descendants: [most influential follow-ups]
bridge_connections:
  - connects_to: "[other anchor]"
    via_paper: "[paper title]"

Phase 3: Integrated Output

Objective: Present unified view of concepts + papers

Output Formats:

Format A: Enriched Concept Cards

# [Concept Name]

**Difficulty:** Intermediate
**Domain:** [domain]

## Definition
[concept definition from domain-vocab]

## Research Origin
- **Seminal Paper:** [title] ([year])
- **Key Authors:** [authors]
- **Core Contribution:** [one sentence]

## Evolution
- **Preceded by:** [parent concepts/papers]
- **Led to:** [descendant concepts/papers]

## Practical Context
[from domain-vocab Phase 4]

## Related Papers
- [paper 1]
- [paper 2]

Format B: Obsidian Vault Structure

{domain}-deep-dive/
├── concepts/
│   ├── {concept-1}.md
│   ├── {concept-2}.md
│   └── ...
├── papers/
│   ├── {paper-id-1}.md
│   ├── {paper-id-2}.md
│   └── ...
├── {domain}-concepts.canvas    # Concept relationship map
├── {domain}-papers.canvas      # Citation tree visualization
└── {domain}-integrated.canvas  # Combined view

Format C: Summary Report

# [Domain] Deep Dive Summary

## Core Concepts (by difficulty)
### Entry Level
- [concepts...]

### Intermediate
- [concepts...]

### Advanced
- [concepts...]

## Research Landscape
### Foundational Papers
| Concept | Seminal Paper | Year | Impact |
|---------|---------------|------|--------|
| ... | ... | ... | ... |

### Evolution Timeline
[Mermaid timeline diagram]

### Key Researchers
- [researcher]: [contributions]

## Learning Path Recommendation
1. Start with: [entry concepts]
2. Read: [foundational paper]
3. Progress to: [intermediate concepts]
4. Deep dive: [advanced + recent papers]

Decision Rules

Anchor Concept Selection

Criterion	Weight	Example
Has clear seminal paper	High	"Attention" → Bahdanau 2014
High interconnectedness	High	Appears in many concept relationships
User interest signal	High	User asked about this specifically
Intermediate+ difficulty	Medium	Entry-level often too broad
Recent research activity	Medium	Active area = more papers to explore

Depth Calibration

User Signal	Action
"quick", "overview"	L1-L2, 3 anchors, shallow trees
Default	L2, 5 anchors, 2-level trees
"deep", "thorough"	L2-L3, 8 anchors, 3-level trees
"researcher level"	L3-L4, all concepts as anchors

Error Handling

Situation	Recovery
No papers found for concept	Use broader search terms, or mark as "practice-origin" (not academic)
Concept too broad (1000+ papers)	Ask user to narrow scope
API rate limits	Queue requests, notify user of delay
Disconnected concepts	Note as "emerging" or "cross-domain import"

Integration Points

With domain-vocab

• Uses: Phase 0-4 (token priming through practical context)
• Extends: Adds paper provenance to each concept

With paper-flow

• Uses: Phase 1-3 (discovery, network extraction, tree analysis)
• Focuses: Only on anchor concepts (not exhaustive field mapping)

Example Session

Input: "Deep dive into Reinforcement Learning"

Phase 1 Output (domain-vocab):

Domain: Reinforcement Learning
Depth: L2 (Practitioner)

Concepts (25):
- Entry: reward, state, action, policy, value function
- Intermediate: Q-learning, policy gradient, actor-critic, exploration-exploitation
- Advanced: PPO, SAC, model-based RL, offline RL

Anchors (5):
1. Q-learning (tabular foundation)
2. Policy Gradient (direct optimization)
3. Actor-Critic (hybrid approach)
4. Deep Q-Network (deep RL era)
5. PPO (practical standard)

Phase 2 Output (paper-flow):

Anchor: Deep Q-Network

Root Paper:
- "Playing Atari with Deep RL" - Mnih et al. 2013
- DeepMind, introduced DQN

Citation Tree:
- Parents: Q-learning (Watkins 1989), CNN for vision
- Children: Double DQN, Dueling DQN, Rainbow
- Bridge: Connects to "Actor-Critic" via A3C (2016)

Phase 3 Output:

• 25 enriched concept cards (5 with full paper genealogy)
• rl-integrated.canvas showing concept-paper connections
• Summary report with learning path

Future Extensions

/1d1s:trace [concept]

Trace a single concept's paper lineage in detail.

/1d1s:frontier [domain]

Focus on L4 (cutting-edge) with latest papers only.

Notes

• Phase 1 runs first, Phase 2 depends on anchor selection
• User can adjust anchors before Phase 2 starts
• Output format can be changed mid-session
• Progress is resumable (concepts extracted → papers pending)