Measured, not promised.

Project measurements from October 2026 runs against real case text. Read them with the caveats — a small holdout bounds the false-clean rate; it does not certify it. This page changes only when new dated runs exist; undated claims live nowhere on this site.

False cleans: 0 observed on 17 (bound ~18%)

A false clean — a bad item marked verified — is the top risk everywhere in this product, so it is the headline metric: 0 false cleans on 17 bad holdout items (rule-of-three upper bound about 18%, reported as a bound, never a rate). The 6 dev items tuned the matcher and count as no evidence. The holdout spans two sanctions matters plus a fabricated statute, so keep diversifying before quoting.

Extraction: 18 of 18

18 of 18 planted citations extracted across 6 synthetic briefs. Extraction is the floor everything stands on: a cite never found is a cite never checked, so misses here would void every number above it.

Identity coverage: ~37.5%

Identity comparison judged on ~37.5% of full citations where a usable party name is available — the rest get existence-only checking, stated as such, never a silent pass. A system that silently passed the other 62.5% would look more accurate and be less honest.

How to read this page

Small samples bound risk; they do not certify safety. The bias everywhere is to flag rather than wave through, upstream failure degrades to "could not verify" rather than "not found," and every filing still needs attorney review under Rule 11. User verdicts, unfiltered — including the 1-star reviews — live on the reviews wall; worked examples with full evidence live in the gallery.