explain

Explains how prompt would be decided in scope, without changing anything.

A read-only companion to lookup for tuning. It embeds the prompt (one Embedder.embed call) and pulls the same nearest candidates, then reports each one's similarity and every guard's verdict — not stopping at the first rejection, the way a real lookup does. It does not touch stats, does not mark any entry recently-used, and does not run the Verifier: a diagnostic that moved the counters you are reading, or spent a model call, would defeat its own purpose.

Reach for it when a hit you expected did not happen. CacheExplanation.decision says whether the threshold or a guard stood in the way, and CandidateTrace.guardVerdicts says which guard and why — across every candidate, so a usable second-nearest entry is visible too.