Skip to main content
Every project starts safe. A baseline model represents your game as it is: control-group players get your default actions, and the experiment has a control worth trusting. Learned models train on the decisions and valued events your game sends. They take traffic only after passing readiness gates: enough decisions, events, and players observed, against thresholds set by policy. Until a model is ready, it cannot be activated; the dashboard shows each unmet requirement with its observed and required counts, and what to do next. Traffic then shifts deliberately, not automatically to 100%. The control group persists so results stay measurable, which is why your code always handles the baseline case: a null Action is the control arm, not an error. The numbers you read in the dashboard carry the same discipline. A metric without enough data behind it is annotated as not yet meaningful rather than presented as a result.