YOU SLEEP. THE AI EVOLVES. 3,500 EXPERIMENTS EVERY NIGHT.
Every night between 1:30am and 6:30am, 11 autonomous engines evolve Strategy Arena. No human touches the code. No external API bill. Everything runs on the VPS.
The 11 engines: Darwin breeds new trading strategies (crossover + elitism + multi-criteria fitness). Leviathan calibrates its 9-Layer Ensemble fusion brain. Portfolio finds the optimal Markowitz mix. Invictus strengthens the immune system on 12 real crashes (COVID -60%, Luna -52%, FTX -23%). Chimera sharpens pattern detection (50 patterns). Hydra tunes ML model weights by regime. Nutrition — the filter that decides what the brain eats — evolves itself. PromptForge tests prompt variants via Claude. Collab Prompts: each AI mutates its own system prompt. Arena Strategies: 1,000 experiments/night on ArenaScript. Meta-Harness: optimizes the optimizer (Darwin hyperparams).
Inspired by Andrej Karpathy's autoresearch. Read more: How we adapted Karpathy's loop · The Living System · Arena Brain
AutoResearch's nightly loop (accumulate → think → act → learn) was validated conceptually on the Dragon Labyrinth Benchmark. Key finding from 14,580 measured trials: human intuition is precompiled compute. An expert who decides in 1 second has amortized millions of rollouts over years of practice. AutoResearch is Strategy Arena's mechanical version of this: 11 engines run every night, mining patterns, retraining models, promoting winners, retiring losers — so the arena wakes up each morning with fresher, better-calibrated priors. Precompilation at night, execution at day — the cognitive shortcut of expérience, reproduced algorithmically.