Run bounded autoresearch

Require a research question, approved datasets, primary metric, hard run and time budgets, output location, and authorization to execute experiments. Use design-experiment to define a falsifiable comparison, fair baseline, seeds, controls, and stopping criteria before spending compute.

Continue in the current Agent: formulate hypotheses, edit configs, interpret evidence, and choose the next experiment with native reasoning and tools. A rendered task, CLI call, or second Agent is not required to begin this work.

For durable cross-experiment budgets, reuse a supplied round or call benchmark.infra.api.create_round with the agreed limits. Attach that round to execution so limits are enforced across sweeps. An optional task template can supply initial defaults through prepare_task; it does not own the research loop. Record only material decisions and evidence references in the round. Native plans and conversation need not be copied into another ledger.

Resolve and preflight every matrix through the library; CLI inspection is an optional adapter. Start with the cheapest experiment that can reject the hypothesis, then use run-experiment with the round attached. Preserve resolved configs, seeds, environments, raw outputs, and failures; never overwrite a costly run. Route failures through diagnose-experiment and compatible results through analyze-results.

When an iteration budget is declared, the Agent defines each research iteration and calls claim_iteration(round_id, operation="stable-iteration-id") once before its work. One iteration may contain several matrices; preparation and resume do not consume iterations. Reusing the same operation ID is idempotent. The optional CLI adapter is tsf research iteration <round-id> --operation <id>. Private reasoning is not automatically metered.

Change one experimental factor per iteration unless an interaction is the stated question. Continue only when the previous evidence justifies the next run. Stop on budget exhaustion, repeated infrastructure failure, invalid comparison, no measurable progress, or a conclusion strong enough for the declared acceptance criterion.

Return the hypothesis, run ledger, compatible metrics with uncertainty, failed or excluded cells, limitations, and a stop/continue recommendation. Repository model code, external publication, and additional task dispatch remain out of scope unless separately authorized.