The Simplicity Criterion tells your agent how to weigh code complexity against metric improvement. The rule from Karpathy's program.md: "All else being equal, simpler is better."
Here's the decision matrix:
- Improvement + simpler code: keep. Best outcome.
- No improvement + simpler code: keep. Simplification win.
- improvement + lines of hacky code: discard. Not worth it.
- + improvement + added complexity: keep. Justified.
Without this directive, your agent will pile on complexity for tiny gains. After experiments, train.py becomes unreadable.