Run measured research

Use research when progress is measured by experiments, such as reducing parser runtime while preserving its output. A research area can contain ordinary supporting tasks and hypothesis tasks. Use investigate for a single question that does not need an ongoing experiment record.

Examples

With the area and research policy ready, paste into chat:

Use zdev to run serial research in area parsing within its brief
for 20 work attempts.

Use zdev to resume the existing research run in area parsing.

Prepare the area

A research run needs an open area with a brief and configured research policy. The policy names the metric, direction, evaluator, measured baseline, correctness checks, permitted changes, and attempt limits. For a parsing area, you could ask:

Use zdev to inspect the research settings for area parsing and explain which
measurements and correctness checks a run will use.

Inspecting or recording settings and hypothesis evidence does not run an evaluator. The research guide explains setup and execution; the research record contract defines the required settings and evidence. Supplied samples must be real measurements.

Authorize a run

With the brief and policy ready, explicitly authorize a run as shown above.

The agent can generate scoped hypotheses and supporting tasks, implement candidates, run configured measurements, and retain the resulting lessons. Keeping improving code requires both the measurement policy and an independent acceptance review. Rejected candidates are retained for inspection while their owned changes are restored; the completed hypothesis still records what was learned.

Twenty work attempts is the default run allowance, not twenty completed experiments. Preparation, measurements, repair, and review consume attempts. You can also specify a time limit, command limit, or target. A run can generate more scoped work when its queue empties; an ordinary loop stops there.

The report names accepted changes and commits, experimental decisions, remaining allowance, and the stop reason. Reaching a budget ends new work; it does not establish that the area’s optimization objective is achieved.

Resume or run in parallel

After an interruption, ask to resume the existing run.

Recovery retains the original run limits and consumed work, and checks pending measurements before launching anything else. With a supporting integration, you can explicitly request rolling parallel research with separate worker and measurement limits. Codex supports this route; check your installed harness integration before assuming the same continuation controls elsewhere.

See the research guide for run controls and the local experiment lifecycle for a single hypothesis. Use recover for interrupted ordinary work.