autoresearch

Iterative code refinement is automated by running a stateful mission loop that continues past non-passing iterations until a bounded max-runtime stop condition is reached. It persists machine-readable evaluation JSON and markdown decision logs for each attempt to ensure strict evaluator-guided improvements.

The evaluator output must be a JSON object containing a boolean `pass` field and an optional numeric `score`. This strict JSON contract ensures decisions are machine-checkable and allows the improvement loop to persist accurate evaluation logs.

Stateful specification tuning is run by executing a single mission directory with an explicit max runtime budget. The loop evaluates each iteration against a JSON contract, halting only when the evaluator passes or the bounded max-runtime stop condition triggers.

Yes, cron integration can be used to schedule repeated experiment evaluation. The loop persists per-iteration evaluation JSON and markdown decision logs under `.omc/autoresearch/`, allowing scheduled runs to maintain auditability and reuse prior state artifacts.

An iterative improvement loop continues running past failed evaluations because it is designed to stop only when max-runtime or another explicit terminal condition is reached. It enforces continuous refinement until the bounded stop behavior triggers a halt.