ARC puzzle visualizer for verifying test integrity and comparing AI responses
- Validates ARC quiz files against full puzzle datasets
- Visualizes AI model outputs in human-readable color grids with numbers
- Highlights mismatched grids between quiz/full and AI/actual answers
- Fully client-side – works offline or on GitHub Pages
- Visit: https://g4te-keep3r.github.io/visARC/
- Upload your ARC quiz JSON and/or full puzzle JSON files
- (Optional) Paste AI-generated answer JSON to compare against correct solution
Under sample_puzzles are 3 random selections from ARC-AGI. Each is included in full and quiz form. Also included is a puzzle of my creation that has produced 11 different wrong answers acros ChatGPT and Claude with various models (GPT-5 and Opus 4.1 now included).
Open index.html in your browser. No setup needed.
Grids are rendered with ARC-standard colors and also display the cell values for easy debugging or annotation comparison.
This example shows:
- Quiz puzzle (top) and Full puzzle with test output (bottom)
- Submitted AI answer (bottom-right)
- The red border highlights a mismatch between AI answer and correct answer
If any training examples differ between quiz and full files (aside from the quiz lacking a test output), they’ll also be shown with red outlines.
- Add examples from https://arcprize.org/blog/arc-agi-2-technical-report - prob need to create jsons by hand
- Fuller explaination on the red grid with example of file integrity error
- This falls under UX improvements, but scaling NEEDS work