The full readout, and the peek row that records having looked at it.

THE PEEK IS PERSISTED. The decision threshold this engine ships is calibrated for weekly looks and this endpoint is unthrottled, so the log is the only thing that can ever say whether that caveat held — and it is what makes the guardrail persistence rule computable at all. A look that could not produce a number (a draft, or an experiment with no control group) comes back readable: false and is not logged.

Every proportion in the response is a fraction in 0..1: each metric value, p_better, p_equivalent, and both lift intervals — where 0.25 means +25%, not +0.25%.