Percentage rollouts need durable exposure and outcome correlation

General lesson: A deterministic percentage bucket is not enough to measure feature lift. Log assignment/exposure for both treatment and control, persist or version the policy used at exposure time, and propagate experiment identifiers to downstream outcome events. Treatment-only operational logs can diagnose failures, but they cannot provide an unbiased denominator or reconstruct controls after the rollout percentage changes.

Applicability: percentage-gated UX experiments whose outcomes occur later in upload, OCR, submission, or verification systems. Limit: deterministic bucket recomputation is adequate only when the exact cohort version and complete configuration history are durable and all eligible subjects can be reconstructed; even then, explicit exposure is safer.

Applies when: Evaluating whether a feature-flag rollout can prove product improvement rather than only technical health. Keywords: feature flag experiment exposure control denominator cohort downstream metrics