Step 5Training teaches it to bet
🧪
Synthetic data only + RLCD Reinforcement Learning for Calibrated Decisions
It doesn't teach it that "ugly" is offensive: it already knows.

1. Answer only by choosing

The comment seems offensive because… ✗ ✏️
0 · Polite or neutral
1 · Critical but respectful
2 · Combative or sarcastic
✓3 · Insults, hate or attacks

2. Honest probabilities

If it says 80%
✓ ✓ ✗ ✓ ✓ ✓ ✓ ✗ ✓ ✓
…it should be right about 8 times out of 10.