Only a holdout tells you what would have happened anyway.

Credit and causation are different columns

An always-on brand campaign can be credited with a large share of conversions while a geo holdout records most of those conversions arriving in the held-out regions anyway. The interface never promotes an attributed figure into a causal claim. Attributed and incremental carry distinct labels wherever either appears.

The test design is fixed first

The experiment service records the hypothesis, the holdout split, the minimum duration and the metric before any budget moves. Readings taken before the duration completes are shown as incomplete rather than as a finding. Stopping early is a decision you take and sign, not something the system does to flatter a result.

Results feed the Constitution, not a headline

A measured incremental ratio becomes evidence a budget proposal is allowed to cite. Where the Campaign Constitution sets a minimum evidence level, a channel with no recent experiment cannot be scaled past its cap on model credit alone. The experiment, its design and its interval are attached to the action record.