Training Pipeline

Stages

1. Knowledge Baking — Commands, recipes, enchantments as training examples
2. API Distillation — Frontier models generate gold-standard responses
3. RCON Validation — Commands executed on real server
4. Self-Play — Model generates edge cases, learns from failures

Data: 3,500+ examples

Category	Count
Command syntax	107
Crafting recipes	176
Enchantments	60
Entities/mobs	60
Gamerules + safety	224
Tool-calling	1,159
API-distilled	344
Bot interactions	448+