Where should the next dollar go?

Seven tiers, from a regular expression to a human, measured on your own records.
The answer is rarely “buy the bigger model”, and every figure can be verified by you.

git clone https://github.com/ArslaneSempai-ui/cascade-routing node src/premiere-reponse.mjs The conclusion, from the sealed records. Before npm install, under a second.

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.You are here · 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.Open Screening

one chip: ten points of measured accuracythe published pick: name goes to the large readerthe empty row: the human tier, never sampled

finding 0117.7 points apart

Both numbers are true. Only one leaves your desk.

94.4%The dashboard mean.

76.7%All five fields, correct. The number the desk works from.

The measured relief, state 02: The cheaper routing

name changes reader: the file-aimed pickthe published pick it replacesa pick both routings share

finding 023.5× cheaper

Cost separates the two routings. Accuracy does not.

$191What the published routing costs, per 100,000 documents.

$54The routing aimed at the file. No file comes out worse.

The measured relief, state 03: Silence over a guess

an emptied cell: silence instead of a wrong value85 wrong values removed, 12 right lost

finding 03after abstention

A blank gets read again. A wrong value gets filed.

30%When every value is delivered, right or wrong.

62.3%When the tool stays silent instead of guessing.

The measured relief, state 04: What we withhold

run twice: every count identical, to the digitnothing turns green: the durations moved, withheld

finding 04withheld

Every count held across two passes. Every duration moved.

identicalEvery count, run twice, to the digit.

16–60%Every duration moved. So durations stay withheld.

The measured relief, state 05: The engagement

every stack green: 16,807 routings crossedstill empty: the human tier, never sampled

finding 05the full span

Every routing enumerated. One report you can argue with.

16,807Every routing tried, end to end. Not a sample.

120 filesHeld out and frozen, 649 tests counted on your machine.

Pick any cell, read what your routing costs.

Accuracy of each tier on each field, measured on held-out records
field	rules	small	large	gen-0.6b	gen-4b	gen-8b	human*
name	0%	46.6%	96.6%	80.8%	89.2%	91.7%	85%
birth	100%	97.9%	100%	87.5%	99.2%	100%	85%
document	79.7%	57.7%	64.4%	70%	79.2%	83.3%	85%
country	100%	100%	100%	83.3%	100%	100%	85%
address	0%	43%	32.8%	75%	95.8%	82.5%	85%

Measured on 1,000 held-out records for rules, small, large; 120 for the generative tiers. *The human tier is assumed at 85% until you measure it: npm run measure:humans grades your own reviewers. Green cells mark the published routing.

Cascade · RoutingOpen the live instrument Every field at every tier, accuracy and cost read live from the sealed record, and a budget slider that chooses the way the tool does.