modbender Quality Filter Skill
This skill provides systematic quality filtering for academic papers with a rigorous scoring system and complete audit trail for research workflows.
Capabilities
- Relevance Scoring: Evaluates paper relevance based on title and abstract keywords
- Quality Assessment: Assesses technical quality and experimental rigor
- Comprehensive Logging: Maintains detailed records of all filtering decisions
- Manual Recall Support: Preserves filtered papers for potential human review
- Local File Storage: Saves all results to organized directory structure
Filtering Criteria
Relevance Scoring (Max 3 points)
- Strong Match (+3): Title contains "music" or "song" keywords
- Medium Match (+2): Title contains "audio" + "generation"
- Weak Match (+1): Title has weak but related keywords
- Negative Scoring: Abstract verification can subtract points (-1 to -3)
Quality Assessment (Max 3 points)
- High Quality (+3): Complete experiments, multiple baselines, strong results
- Medium Quality (+2): Experiments present but limited baseline comparison
- Low Quality (+1): Limited technical contribution or incomplete evaluation
Pass Threshold
- Minimum Score: 6/10 points required to pass filtering
- Strong Relevance Override: Papers with clear "music/song generation" focus may pass with lower scores
Workflow Integration
This skill integrates with the broader research workflow:
- Input: Raw paper list from arXiv search
- Processing: Applies scoring system to each paper
- Output: Categorizes papers as "passed" or "filtered"
- Audit Trail: Maintains complete record for manual recall
Output Format
Local File Storage
- Main Log:
research/{domain}/quality_filtering/quality_filtering_log.md - Append Mode: All results appended to single comprehensive file
- Directory Structure: Automatically created if missing
Log Structure
Each filtering session includes:
- Session Header: Date, domain, search parameters
- Scoring Standards: Detailed criteria used
- Individual Paper Results: Title, authors, score breakdown, decision
- Summary Statistics: Pass/fail counts, score distribution
- Manual Recall Section: List of filtered papers available for human review
Usage Examples
# Filter music generation papers
quality_filter --domain "music_generation" --papers "[paper_list]" --date "2026-02-28"
# Filter with custom threshold
quality_filter --domain "speech_audio" --threshold 5 --papers "[paper_list]"
Files Created
research/{domain}/quality_filtering/quality_filtering_log.md(append mode)- Directory structure automatically created if missing
Audit Trail Requirements
All filtering decisions must include:
- Complete score breakdown (relevance + quality components)
- Clear pass/fail rationale
- Preservation of filtered papers for manual recall
- Timestamp and session context