Get Streak Programme

Is the streak ladder selecting anybody, and can we measure their streaks at all?

This is the shadow gate, not the causal readout. The treated-versus-control comparison for any streak agent is GET /agents/{agent_id}/holdout, which runs on the shared holdout service: one grouped query, one maturity clause, one outcome for both arms. Recomputing arms here would be a second implementation of the same split, and any drift between the two would read as a real effect rather than as a bug. So this answers only the two questions that are specific to this programme and that the rollout is blocked on.

Coverage. customer_events is a FILTERED store: the extractor writes only events classified core, minus the workspace's machine-event names, and only for workspaces that opted into extraction. A server-emitted streak event can be excluded from it entirely, so "the event is there" is a measurement rather than an assumption. The split below is how many people the publisher could measure from events against how many fell back to the person property, which is unreliable at the exact moment the loss ladder fires.

Volume. Enrolment by moment, from the currently published state, plus how often streaks actually break and get rebuilt in this workspace. The break figure is descriptive and explicitly NOT causal: it counts everybody, messaged or not, and exists to size the programme rather than to credit it.