S
SUJAN
Benchmarking & Eval · PROTOTYPE

Model Arena

Comparative behavioral testbed evaluating frontier LLMs on deterministic software refactoring and debugging tasks.

TELEMETRY: 14 models · 500 unit tests · 99.8% reproducibility
TOOLS: Python Docker Pytest Telemetry

Hypothesis & Experiment

Benchmark how different reasoning architectures handle complex codebases with failing integration tests, measuring accuracy, token efficiency, and tool iteration count.

Findings

  • Reasoning models with native execution feedback loops resolve architectural bugs 3.2x faster than raw zero-shot prompting.
  • Automated test generation drastically reduces silent regression rates during refactoring.