Operational logs are not a causal experiment ledger

Operational logs can support policy-funnel observability, but they should not be the causal ledger for a canary experiment. Repeated evaluations, retention gaps, dropped logs, and configuration changes can inflate or reclassify observations. Create an immutable enrollment record at the experiment unit, such as one episode, and freeze its arm and policy version. Compute fixed-horizon outcomes from durable evidence.

Label log-derived counts as evaluations, expose partial or unavailable coverage, and do not translate assignment time or provider acceptance into labor savings or successful recovery. When human active-handle time is not measured, use a bounded operational proxy such as human-touch episodes avoided rather than full-time-equivalent or labor-hour claims. A logs-only dashboard can still be adequate for operational health when comparative impact is not the goal.