Which matcher suffices, at which threshold?

Seven name matchers, from strict equality to a multilingual embedding, swept across fifty‑one thresholds.
Recall and false alerts carry their intervals, and every figure can be verified by you.

git clone https://github.com/ArslaneSempai-ui/cascade-screening npm ci --ignore-scripts npm run measure:yours -- --alerts=your-alerts.csv Your alert history, measured on your machine. Nothing of yours goes up.

Cascade · two instruments, one method Routing · extraction

Where should the next dollar go?Seven tiers, from a regular expression to a human, measured on sealed records.Open Routing

Screening · name matching

Which matcher suffices, at which threshold?Seven matchers, from strict equality to a multilingual embedding, measured on your own alert history.You are here · Screening

The sieve tower, state 01: Exact never alarms. It also never finds.

the strict sieve, at the top: a few hits, no noisethe loosest sieve, at the bottom: every hit is still there

finding 014 hits in 5 missed

Exact match after normalisation raises no false alert on our written pairs, and that safety is the trap: it misses four true hits in five that your reviewers already confirmed.

21.7%recall on the 60 written match pairs, interval [13–34]: any threshold, exact is flat

0%false alerts on the 60 written non-match pairs, interval [0–6]

The sieve tower, state 02: Loose matching finds everyone. And everyone else.

the loose sieve: nothing escapes, nothing is filteredthe grey ones are false alerts, nearly one per hit

finding 02every hit, drowned

Drop the threshold and nothing escapes: every true hit is caught. So is most of everything unrelated, and every false alert is analyst time you pay for.

100%recall at the loosest threshold of the sweep, interval [94–100], 60 match pairs

85%false alerts at the same threshold, interval [74–92], 60 non-match pairs

The sieve tower, state 03: The frontier lives in the middle.

the ruby frame: the sieve the frontier retainswhat it still lets through lies at the foot

finding 03the frontier cell

Hold the lower bound of measured recall at 90 percent, the tool's own default rule, and the sweep of every tier and threshold, the embedding included, settles on one cell. Even there, three alerts in four are false.

98.3%recall of the retained cell, jaro‑winkler at 0.56, interval [91–100]

76.7%false alerts at that same cell, interval [65–86]: the price of the floor

The sieve tower, state 04: Synthetic is declared. And never merged.

written pairs: the measured halffabricated variants: labelled, apart

finding 04never blended

The same cell, two provenances. On fabricated variants the false-alert rate reads far friendlier than on written pairs: blended, the synthetic half would flatter the tool. It is measured apart, labelled synthetic, and stays apart.

76.7%false alerts at the frontier cell, measured on the 60 written pairs, interval [65–86]

17.2%the same cell on 360 fabricated variants, interval [14–21]: kept apart precisely because it flatters

The sieve tower, state 05: Your history, your frontier.

our pairs end where your export beginsmeasure:yours, next to your file: uncoloured until measured

finding 05yours replace both

Every pair behind these findings is ours: written and declared, or fabricated and labelled. One command replays the whole sweep on your own alert history, on your machine. Nothing leaves it.

60written pairs behind the measured half, declared nature by nature

360fabricated variants behind the synthetic half, labelled and kept apart

Pick any cell, read what your threshold costs.

Recall over false alerts of each matcher at each threshold, on the written pairs
matcher	0.50	0.56	0.60	0.70	0.80	0.90	1.00
exact	22% 0% fa	22% 0% fa	22% 0% fa	22% 0% fa	22% 0% fa	22% 0% fa	22% 0% fa
tokens	52% 3% fa	52% 0% fa	52% 0% fa	48% 0% fa	48% 0% fa	48% 0% fa	48% 0% fa
jaro-winkler	100% 85% fa	98% 77% fa	97% 75% fa	88% 70% fa	72% 62% fa	62% 30% fa	22% 0% fa
damerau	80% 70% fa	78% 63% fa	77% 62% fa	70% 35% fa	60% 20% fa	47% 18% fa	22% 0% fa
phonetic	87% 73% fa	80% 62% fa	78% 53% fa	73% 35% fa	67% 23% fa	47% 7% fa	42% 5% fa
ngrams	80% 50% fa	78% 32% fa	70% 23% fa	58% 18% fa	42% 2% fa	32% 0% fa	22% 0% fa
embed	100% 100% fa	100% 100% fa	100% 100% fa	100% 100% fa	100% 100% fa	100% 78% fa	10% 0% fa

Measured on the 60 written match pairs and 60 hard negatives: recall on top, false alerts below. The ruby cell is the tool’s frontier under its default rule, jaro-winkler at 0.56. The synthetic variants stay apart, on the instrument.

Cascade · ScreeningOpen the live instrument Every matcher at every threshold, recall and false alerts with their intervals, live from the sealed record, and the tool's own selection rule under your recall floor.

Cascade Screening, on film.

The five findings, explained

The ruby Cascade robot, palms up, projecting the two rates of the frontier cell: recall on the left, false alerts on the right.

hosted on YouTube · link pending upload

The appendices your reviewers will ask for.

Method & what is measured What the frontier reads, and what it refuses.

Security & data handling The lists, the seal, and what never leaves.

Terms of engagement What the grant allows, for how long, and what a client buys.

Privacy No data is collected. Written down, and verifiable.

Accessibility Usable by keyboard, by screen reader, and with motion turned off.

Pricing, in figures Contact The fine print The repository, public