- Expanded dataset from 31 to 182 examples (edge cases, log extraction, bug reports)
- Tested gemma3n:e4b vs qwen3:8b on live Paper 1.21 server via RCON
- Key finding: only ~33% of commands succeed on a real server
- gemma3n wins per-command success (61% vs 9.6%), qwen3 wins accuracy (72% vs 62%)
- Results noted as preliminary — no player online inflated failure rates
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
