Karpathy Autoresearch Protocol

You sleep. Our agents evolve 3,500 strategies per night. Meta-Harness optimizes the optimizer.

Welcome to the lab where the strategies that <strong>beat the Arena</strong> are born. Every night, 7 autonomous engines mutate, test, and keep only what works.
Inspired by Andrej Karpathy
×

Every night, <strong>7 AIs test 3,000 mutations</strong> of trading strategies and keep only the best. Natural selection, but for trading.

Mutate
Test
Keep/Discard
Repeat
-
Experiments
-
Improvements
8
Active Engines
-
Best Score /100
LIVE EVOLUTION FEED — DARWIN ENGINE v2.0

Nightly Schedule

Every night: mutate &rarr; test &rarr; keep/discard &rarr; repeat

Hall of Fame

Loading...

Cross-Engine Fitness

Loading...

Nightly Run Log

Loading...

Living Wiki — Lessons

Loading...

The Karpathy Method

1

Mutate

Take the best parameters. Apply small random mutations. Like DNA — most are neutral, rarely one is beneficial.

2

Test

Backtest the mutation on real data. Budget: <code>5 sec</code> per experiment. ~300 experiments per engine per night.

3

Keep or Discard

If <code>universal fitness</code> (0-100) improves, keep. Otherwise discard. Only improvements survive.

4

Repeat

7 independent tracks, every night. Over weeks, converges to optimal. Zero human intervention.

5

Meta-Harness

The agent that optimizes Darwin itself. Mutation rate, crossover ratio, fitness weights — all auto-tuned. <strong style='color:#ef4444'>Darwin evolves strategies. Meta-Harness evolves Darwin.</strong>

Explore Knowledge Graph Nerve Center

How AI Strategy Evolution Works

Strategy Arena's Evolution Lab uses the Karpathy autoresearch pattern to evolve AI trading strategies autonomously. Every night, 7 independent research engines — Darwin, Leviathan, Chimera, Invictus, Hydra, Portfolio, and PromptForge — run thousands of experiments on real BTC market data. Each engine mutates strategy parameters, backtests them against historical prices, and keeps only improvements. This process mirrors biological natural selection: strategies compete, the fittest survive, and weak mutations are discarded. Over weeks, this produces trading algorithms that no human could design manually — optimized across Sharpe ratio, win rate, drawdown, and profit factor simultaneously. All results feed into a Living Wiki that accumulates knowledge across runs, ensuring the system never forgets what works.

Rejoindre le canal 💬 Feedback ☕ Support Us
The terminal below shows live mutations happening right now ×