autoresearch

An iterative improvement loop for machine learning missions iteratively refines models based on evaluation results while adhering to strict evaluator contracts. This approach ensures structured JSON output and durable experiment logging for persistent review.

A strict evaluator contract for structured JSON output mandates that evaluation results include a required `pass` boolean and an optional `score` field. This enforces rigorous, evaluator-driven iterative improvement by standardizing how evaluation metrics are logged and reviewed.

To log experiment decisions and evaluation results for persistent review, the system stores evaluation JSON and markdown decision logs in durable storage. This allows for structured, long-term analysis of evaluator-driven iterative improvements across multiple runs.

Yes, you can use a stateful improvement loop without external dependencies. The system operates with no listed dependencies, relying solely on its internal scripts and assets to manage the bounded, evaluator-driven iterative enhancement process.

The best way to run an evaluator-driven mission optimization is to execute the process with a specified mission directory. This triggers the stateful loop to apply strict evaluator contracts, log results in JSON format, and store markdown decision logs for review.

You should not use a stateful loop for mission optimization if your task requires unbounded iterations or lacks clear pass/fail criteria. The system is specifically designed for bounded, evaluator-driven iterative improvement with strict JSON output contracts.