Autoresearch: Autonomous Improvement Loop

An AI-driven methodology where you iteratively modify a single target, evaluate changes against an objective scoring function, and compound improvements. Based on Karpathy's autoresearch pattern, proven on 103 experiments achieving 7.6x score improvement.

When to Use

• Optimizing trading strategies against a backtest harness
• Improving code quality metrics (test pass rate, coverage, performance)
• Tuning configuration parameters against benchmarks
• Prompt engineering with measurable evaluation
• Any domain where you have: a mutable target + an objective score

The 5 Pillars

1. Single Mutable Surface

Only ONE file (or narrow set of files) can change per experiment. Everything else is frozen. This constrains the search space and makes experiments reproducible.

2. Immutable Evaluation Harness

The test/scoring system NEVER changes during the loop. If you change both the target and the evaluator, you can't tell what improved.