Take the instrument. It answers live.

One tier per field: click any cell, read what your routing costs and how much it gets right. Or slide the budget and let the enumeration decide, the way the tool does: highest accuracy first, cheaper on a tie.

The Cascade robot leaning over the terminal from behind its frame

cascade · live instrument

$ cascade 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

Pick one tier per field; each cell shows measured accuracy and the price of a thousand extractions
field	rules	small	large	gen-0.6b	gen-4b	gen-8b	human*
name
birth
document
country
address

$ cascade optimise --budget

Budget, dollars per hundred thousand documents, logarithmic

slide to read the best routing under your budget

$ cascade verify --sealed

self-check requires JavaScript; the figures above are still the sealed readings.

What this instrument rests on, and what it refuses.

The prices are assumed, and say so: the small and large tiers are billed at assumed per-call rates, the generative tiers at their frozen measured latency times an assumed machine cost, the human tier at an assumed pace and salary. Change the assumptions and the dollars move; the accuracies do not.

The human column is an assumption, never a measurement: 85% on every field, declared in the tool's own source. It is the only displayed figure that was not measured.

Your documents never touch this page: nothing is uploaded, nothing is fetched, and the page's security policy refuses every network call. These are our sealed readings; your records may disagree, and there is one honest way to find out.