autoresearch

To run autonomous experiments, you select a mode like Optimize, define a mutable surface, a truth layer for tests, a numeric verifier, and a metric direction to iterate and track provenance over time.

An autonomous research loop iteratively modifies prompts or configurations, measures results against a numeric verifier, and logs the experiments to establish baselines and track provenance for replayable improvements.

Yes, you need a truth layer consisting of tests or benchmarks to serve as the numeric verifier, ensuring that each iteration of the mutable surface is measured against a clearly defined metric and direction.

The best way to track provenance is using a strict branch and log workflow that records each experiment in a results ledger, establishing baselines and ensuring the entire optimization process remains replayable.

Yes, the Research mode utilizes multi-perspective research patterns like STORM to perform deep analysis, informing subsequent modifications to the mutable surface before validating them with the numeric verifier.

You should avoid this loop if your software artifact lacks a clearly defined mutable surface, a measurable numeric verifier, or a specific metric direction, as these are strictly required for the experiment workflow.