CRUSETRA

Test the routing on our public test set, live.

One model tier per field, which is the size of model that field is sent to. Click any cell to see what that routing costs and how often it gets the field right, or set your budget and let the tool choose, the way it does on your machine: highest accuracy first, lower cost when two tie.

crusetra · live instrument

$ crusetra compose --live

your routing  name:large · birth:rules · document:rules · country:rules · address:gen-4b
cost         $191 /100k docs · assumed prices
accuracy     94.4% per-field mean · no interval
vs published +$0 · +0.0 pt

each dot is one routing of the five fields, priced and scored from our frozen readings. The line through the bright dots is the best trade-off no routing beats · pull the budget line (or the slider below), the best routing under it lights up · hover a frontier dot to read it, click it to compose it

Pick one tier per field, and each cell shows the measured accuracy and the price of a thousand extractions
fieldrulessmalllargegen-0.6bgen-4bgen-8bhuman*
name
birth
document
country
address

$ crusetra optimise --budget

slide to read the best routing under your budget

$ crusetra verify --sealed

self-check requires JavaScript. The figures above are still our frozen readings.

What this instrument rests on, and what it cannot do.

The prices are assumed, and the page labels them as such. The small and large tiers use assumed per-call rates. The generative tiers use their measured latency against an assumed machine cost. The human tier uses an assumed pace and salary. Change the assumptions and the dollars move. The accuracies do not.

The human column is the one exception, an assumption until you measure it. We assume 85% on each field, which the tool declares in its own source. Each other accuracy shown here was measured. npm run measure:humans -- --cases=your-file.csv grades your own reviewers: accuracy per field with intervals, agreement between reviewers, seconds per record. Pass the sealed result to optimise with --humans and the optimizer reads your measurement instead of the assumption.

Your documents never touch this page, because nothing is uploaded and nothing is fetched: the browser's own security rules refuse each network call.

These are our readings. Now test them against yours. The tool clones next to your files and measures them where they sit, so your CSV never moves and the report lands beside it.