Select saved files
Use title/body snapshots from search-docs.js and a
searchlite.local-judgments/1 labels file. Both indexes must contain identical
stored document IDs and title/body text.
Up to 16 MiB per index, 128 documents and 1 MiB of body text; labels up to 256 KiB and 32 queries. Comparison runs in a worker with a 15-second deadline. The already-loaded page works offline.
Choose three files or load a synthetic example.
Try a complete example
Two fictional notes put “quartz” in different fields. Reweighting title and body improves one query and regresses another; the individual drop still fails even when both minimums pass.
What the result means
Every candidate query must meet its declared minimum and have unrounded nDCG@10 at least as high as its baseline. A mean improvement cannot override a query regression. Rankings use the actual supplied indexes, with the same engine and rules as the native CLI.
This is a deterministic gate for supplied labels, not a significance test or proof of general retrieval quality. Matching stored text and SHA-256 hashes do not authenticate postings, configurations, labels or sources. The result does not establish current source freshness. Clear drops application references; it does not promise secure memory erasure.