Autoresearch

Run isolated experiments, measure them consistently, keep verified improvements, and stop on explicit resource or convergence limits.

Session Contract

Resolve the objective, primary metric and direction, benchmark and correctness commands, allowed/off-limits paths, run/runtime/command/cost/regression limits, convergence window, and reporting cadence before the baseline. Infer safe facts from the request and repository; ask only when a missing choice changes the experiment.

Defaults: 20 runs, two hours wall time, 10 minutes per benchmark, five minutes per correctness check, no new paid API spend, and convergence after five consecutive valid runs without a new retained best. Explicit --max-runs and --max-runtime values are hard limits.

Isolation