Our research

Proving a cheaper model got no worse.

Routing to a cheaper model is easy. Deciding it changed nothing, without a human reading every prompt, is the program.

See a chain re-derived

A number and the arithmetic under it.

What makes it hard

Real traffic
Leaderboard scores do not transfer to your own prompts.
No human in the loop
A person grading every prompt is a demo, not a system.
Cheaper than the answer
An evaluation bill that eats the difference proves nothing worth having.
Not proven
Uncertainty resolves downward. It never rounds up into a claim.

Recovea research

Every number the platform shows is one click from the arithmetic behind it, and every receipt re-derives without us.

Read the docs

The words we print beside a number

Measured: Arithmetic over counts the provider itself reported, priced against a version-pinned list. An unreported count is stored as null, never as a guess.
Modeled: A measured observation carried past its window. The label travels with the figure, and a projection is never rounded up into an outcome.
Applied: Something actually changed your traffic. Nothing here is applied: no lever is switched on in any deployed environment.
Verified: Reserved, and unused. No figure on this site carries it, and the platform vocabulary has no value that could hold it.

The price list is frozen

Costs are computed against a published reference list, pinned by version on every receipt. A published version never changes: a correction ships as a new one.

The chain recipe