Held-out evaluation
Scoring a memory with the answer barred from participating — the only way a score means anything once the data is already stored.
Like you're twelve
Asking a friend to recite a poem while they are holding the book open is not a memory test. Close the book and it is.
A memory that has absorbed an item will hand that item back nearly perfectly. That number tells you the write worked, not that the memory generalises.
Precisely
Every self-query against stored data returns a high score by construction, so it carries no information about whether anything was learned. An honest measurement excludes the item being asked about — and, where the store holds near-duplicates, its whole family, since one near-copy answers as well as the original.
The engine exposes this as leave-one-out reads: a read with a single stored engram excluded, or with a family excluded. Most of that family lives in the Rust library rather than on the HTTP surface, so an HTTP client typically holds out a split before writing: keep a portion of the data unwritten, score against that, and report coverage alongside accuracy.
The same discipline applies to every threshold. A gate chosen by judgement is an untested guess; a gate derived from held-out scores has a measured error rate at a measured coverage. Thresholds are valid at the load they were calibrated at, and drift as the pool grows.