Refactor toxicity analysis implementation to be independent of specific
LLM providers. Update configuration and code to use generic terminology
and allow flexibility in choosing language models.
Changes:
- Update environment variable naming for API credentials
- Generalize documentation to support multiple LLM providers
- Improve configuration flexibility for model selection
- Add project documentation files to gitignore
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Added comprehensive project documentation and MIT license. Removed Claude
IDE configuration files from repository tracking.
Documentation added:
- FINDINGS.md: Complete methodology report and research findings
- 159 accounts tracked, 15,190 posts collected (Jan 1 - Mar 30)
- Human review results: 40.4% correct, 59.6% false positives
- AI toxicity detection limitations and recommendations
- OPERATIONS.md: Complete operations and maintenance guide
- Service start/stop procedures
- Database operations and queries
- Configuration options
- Troubleshooting guide
- Data export instructions
License:
- Added MIT License to README.md
- Copyright 2026 Post X Society
- Open source with permissive license
Repository cleanup:
- Added .claude/ to .gitignore
- Removed .claude/settings.local.json from tracking
- Prevents IDE-specific files from being committed
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Enhanced the web interface with comprehensive human review capabilities
for validating AI-flagged toxicity classifications. Added date filtering
and improved data collection to include engagement metrics for mentions.
Features added:
- Human review system with ✓/✗/? status buttons and filtering
- Date range filtering (from/to) for flagged content
- Review status tracking with database migrations
- Engagement metrics collection for mentions (likes, replies, reposts, quotes)
- Interactive review buttons that allow changing classifications
- Review filter to show unreviewed, correct, incorrect, or unsure items
UI improvements:
- Fixed Chart.js CDN URLs (switched to jsdelivr)
- Smart axis scaling for toxicity category charts with dynamic decimal places
- Clickable max toxicity badges linking to filtered content
- Improved mention author display using raw_json fallback
- Sortable table columns with visual indicators
- Review status preserved across pagination and filtering
Bug fixes:
- Commented out problematic account (stephanvanbaarle.bsky.social)
- Fixed filter parameter names (content_type, account_did)
- Fixed threshold boundary issues with 0.001 offset
- Added extra_js block to base template for JavaScript functionality
Database changes:
- Migration 03: Added engagement columns to mentions table
- Migration 04: Added human_reviewed, review_status, reviewed_at columns
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Bluesky post collector with mention tracking
- PostgreSQL database for storage
- OpenAI-based toxicity analysis
- Web UI for viewing and analyzing posts
- Docker compose setup for deployment
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
