Phase 3: Filter

You are executing Phase 3 of the research funnel methodology. Your goal is to apply rapid evaluation criteria to filter 15-30 candidates down to 3-5 high-value papers for deep analysis.

Input

{{#if session}} Session: {{session}} {{else}} Session: Auto-detect from most recent session in ./research/ {{/if}}

Target Selection Count: {{count | default: "5"}}

Instructions

Step 1: Load Previous Phases

1. Load phase-1-define.md for research question context
2. Load phase-2-discover.md for candidate list
3. Verify both exist before proceeding

Step 2: Apply 10-Second Scan

For each candidate from Phase 2, quickly evaluate using these criteria:

Primary Filters (Must Pass)

• Relevance: Directly addresses the research question
• Recency: Within acceptable date range (or seminal work)
• Accessibility: Can actually access the full content

Quality Indicators

• Citation count (if available): High citations suggest impact
• Venue quality: Top conferences/journals score higher
• Author reputation: Known researchers in the field
• Abstract quality: Clear problem statement and contributions

Practical Factors

• Code availability: Implementation to study
• Reproducibility signals: Clear methodology, datasets mentioned
• Length/depth: Appropriate for the type of contribution

Step 3: Score and Rank

Create a scoring matrix:

Criterion	Weight	Score (1-5)
Relevance to question	3x
Technical depth	2x
Novelty/contribution	2x
Clarity/accessibility	1x
Reproducibility	1x
Code/data available	1x

Total Score = Σ(Weight × Score)

Step 4: Select Top Papers

1. Rank all candidates by total score
2. Select top N papers (default: 5)
3. Ensure diversity:
   • Mix of approaches/methods
   • Mix of publication types if appropriate
   • Include at least one recent paper

Step 5: Document Rejections

For each rejected paper, briefly note why:

• "Too tangential - focuses on X not Y"
• "Superseded by newer paper Z"
• "Paywalled with no alternatives"
• "Low quality/predatory venue"

Step 6: Write Output

Create ./research/<session>/phase-3-filter.md:

# Filter Results: <Session>

**Session**: <session-slug>
**Created**: <ISO timestamp>
**Phase**: 3 - Filter
**Based on**: phase-2-discover.md

## Selection Summary

- **Candidates evaluated**: <n>
- **Papers selected**: <n>
- **Papers rejected**: <n>

## Selected Papers

### 1. <Title> ⭐ [Highest Score]

- **ID**: <P#>
- **Score**: <total>/50
- **Why Selected**: <1-2 sentence justification>

| Criterion | Weight | Score | Weighted |
|-----------|--------|-------|----------|
| Relevance | 3x | <n> | <n> |
| Technical depth | 2x | <n> | <n> |
| Novelty | 2x | <n> | <n> |
| Clarity | 1x | <n> | <n> |
| Reproducibility | 1x | <n> | <n> |
| Code available | 1x | <n> | <n> |

**Reading Priority**: <Primary/Secondary>
**Expected Contribution**: <What we hope to learn>

[Repeat for each selected paper...]

## Rejected Papers

### Quick Rejections (Did not pass primary filters)

| ID | Title | Rejection Reason |
|----|-------|------------------|
| <P#> | <Title> | <Reason> |

### Scored but Not Selected

| ID | Title | Score | Why Not Selected |
|----|-------|-------|------------------|
| <P#> | <Title> | <n>/50 | <Reason> |

## Selection Rationale

<Paragraph explaining the overall selection strategy and how these papers complement each other>

## Gaps in Selection

<Any areas of the research question not well covered by selected papers>

## Next Phase

Run `/researcher:deep-dive` to perform three-pass analysis on selected papers.

Execution

1. Load Phase 1 and Phase 2 outputs
2. For each candidate, apply the 10-second scan
3. Score each paper that passes primary filters
4. Select top papers ensuring diversity
5. Document all rejections with reasons
6. Write the output file
7. Summarize selection and suggest running Phase 4

Tips for Effective Filtering

• Trust your initial instincts from the 10-second scan
• Prioritize papers that can teach you something unique
• Don't just pick the highest-cited papers - include innovative newer work
• Consider how papers will work together in synthesis
• It's okay to adjust selection after starting deep-dive if needed