Training Examples (I/O pairs)
1
Programmer
LLM generates
transform(grid)
code
2
Validator
Execute code on training pairs
On failure → LLM fixes code
up to 2×
3
Code Execution
Run validated code on test inputs
4
Fallback
Direct LLM grid prediction
Attempt 1
Attempt 2
Assembly
Combine code output (Attempt 1) + fallback (Attempt 2)
Final Predictions (2 attempts)