Experiments
An Experiment is an approach or hypothesis the agent investigates against your Objective. It carries a confidence score from the reviewing models and can produce several Versions. A draft Experiment has not run yet, and a generated Version has not necessarily been tested.
The Experiments page of a Discovery Run has four views: Canvas (the default, see Canvas), Kanban, List and Archive.
Statuses
On the Kanban board, Experiments move left to right:
- Draft: proposed but not accepted yet. Queue it to accept it, or Dismiss it.
- Queued: accepted and waiting to run.
- In progress: its Versions are being generated and run on your Runner.
- Done: finished and recorded as Validated, Refuted or Inconclusive.
Dismissed Experiments move to the Archive view, where you can restore them.
The outcomes are the run's recorded conclusions under its evaluation method. Validated means the evidence supported the hypothesis, Refuted means it did not, and Inconclusive means the evidence did not settle it. Read them alongside the per-Metric results: Validated on its own does not prove that tests pass or that a gain is statistically reliable.
Working with the board
Search, filter by creator, and sort Experiments by status, confidence, created date or task name. Add your own with New. Select several Experiments to act on them together, for example to Combine them.
In Manual approval mode the run waits once the agent has proposed Experiments. When at least one is queued, the banner says Ready to resume. Click Resume and the agent picks up the queued Experiments, generates their Versions, runs them on your Runner and scores them against your Metrics. It then pauses again for your next round. In Automatic mode it keeps going until the run completes.
Each finished Experiment links to the Versions it produced.