Which matcher, at which threshold. Live.

Every figure on this page comes from the sealed public record of cascade-screening: pairs the repository wrote itself, hard negatives included, measured by its own matchers. No client data exists here, nothing enters and nothing leaves, and the page checks its own arithmetic before your eyes.

cascade screening · the public record, live

$ cascade screen --live

Pick a cell: each shows recall on confirmed matches over false alerts on hard negatives, at that matcher and threshold
matcher \ threshold	0.50	0.60	0.70	0.80	0.85	0.90	0.95	1.00
exact
tokens
jaro-winkler
damerau
phonetic
ngrams
embed

pick a cell: recall over false alerts, with n and its 95 % interval

$ cascade optimise --recall

$ cascade verify --sealed

checking…

What this rests on

The sealed public record. releve-public.json in the repository, fingerprint 74a18ebf30d05515, measured at commit e86a139 on 2026-09-05. The extractor that feeds this page verifies the seal, then recomposes witness cells with the tool’s own interval code, and refuses to emit if a single figure disagrees.
Pairs the repository wrote. The labelled half is authored: true correspondences (transliterations, token order, initials, typos, particles) and hard negatives (siblings, partial homonyms, near-strings that are not the same person). The labels ship with the pairs, debatable ones with their reasons.
Synthetic variants, kept apart. Generated from list-entry names, nature by nature, and never merged with the authored half: the toggle above switches the whole grid, it never blends.
The tool’s selection rule. The slider holds your recall floor at the Wilson lower bound, exactly as npm run optimise does: a point estimate never clears a floor here.

What this refuses

Your data. This page cannot read it: no network requests leave it (connect-src ‘none’), no third-party resource is loaded, and there is no input field to paste a name into.
Bare rates. Every cell carries its n and its 95 % interval; every matcher of the registry is present tonight; an absent one would be a named row, never a guessed column.

Measure your own alert history

The instrument shows our record. Yours is measured at home, by the tool, and nothing about your file leaves your machine.

The three commands, exactly as they run

$ git clone https://github.com/ArslaneSempai-ui/cascade-screening.git

$ npm ci --ignore-scripts

$ npm run measure:yours -- --alerts=your-alerts.csv

the report and the sealed record are written next to your file, and nowhere else