A test queue with holdouts, not a calendar of launches.

Hypotheses stored as objects, not slides

A test is declared before it runs: what is expected to change, on which metric, by when, and what would count as a failure. The declaration is stored with the result, which is why a test can be re-run a quarter later and compared against the original rather than argued about.

Holdouts and geo tests run as tests

Where the question is whether spend changed the outcome, a holdout region or a paused cell is set up with its exposure and control defined in advance. Attribution continues to distribute credit under your chosen rule in parallel, and the two readouts are reported next to each other because they answer different questions.

The model plans; the services compute

Cost per outcome, margin, attributed credit and test significance are produced by deterministic services reading your connected data. The model's job is choosing which of those calls to make, in which order, and turning the output into a readable explanation with the inputs cited.

Losing variants leave the rotation automatically

When a readout date passes and a variant has lost against the declared metric, it is pulled from rotation and the outcome is recorded against the original hypothesis. Whether that happens on its own or needs a person is one of the thresholds in your Campaign Constitution.