Decisions that cannot be replayed

What it looks like. Logs record what was decided, not what the decision saw: no inputs, no retrieved context, no model state, no versions. Sampling is non-deterministic and compliance is checked against a sample.

Why it fails. Without replay you cannot distinguish a correct decision from a lucky one, or an incident from a pattern. Sample-based checks on a stochastic system license retry-until-permitted: the population fails at a rate the sample never sees.

Instead. Make replay-readiness a service objective for high-consequence actions, logging inputs, context, retrieved evidence, and model state. For policy-critical decisions require deterministic output or a post-hoc filter, and track the tail rate rather than the sample rate.