Dual LLM Judge System Launches, Three Kings Crowned

SN66 deployed a dual-judge scoring mechanism using OpenAI GPT-5.4 and Claude Sonnet 4.6 with blinded deliberation, replacing the old similarity-based system. Three new kings were crowned in 24 hours despite increased difficulty. The validator also shipped three critical fixes: closed re-eval exploits by limiting evaluations per repo to one, removed stale-king-hash checks, and corrected dashboard earnings calculations (cut from $547/hr to ~$240/hr per king). A Discord outage briefly affected the dashboard, now recovering.

• •Dual LLM judge live with blinded deliberation; margin 3 and solver minimax-m2.7 unchanged
• •Dashboard earnings corrected: ~44% of gross subnet emission, not 100%
• •PR-based miner entry system requires exact SHA match between on-chain commit and GitHub head
• •Agent timeout: min(max(cursor_elapsed × 2 + 1s, 120s), 600s) hard wall
• •Miners can test locally via validator_harness_v5.py; aim for ≥65% win rate on 20 tasks