Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Verifier

  • Checking AI's work
  • Standard term
3 steps

A verifier is a check that tests one stated property of an AI's output, such as whether required fields are filled in, code passes its tests, or a claim matches its cited source.

A rule written in code, such as "every price is a number," gives the same verdict every time it runs. A property that takes judgment, such as tone, can go to an AI grader instead, whose verdict on the same answer can change from one run to the next.

Where you’ll meet it

Your team's AI tool reads supplier emails and adds rows to the inventory sheet. A check confirms that each row has a product code, a whole-number quantity, and a price. Every row passed last week, and one quantity read 100 where the email said 10.

What to do

Ask whoever runs the tool to add a second check aimed at this failure, one that compares each quantity with the number in its source email and sends any mismatch to you before the row is saved.

Easy to mix up

An eval judges a whole system, such as whether it is ready to release. A verifier acts on one output at a time, deciding whether to keep it, redo it, or reward it.

Check yourself

Your coding agent adds a discount-code feature to your shop, writes tests for it, and reports that every test passes. What does the pass tell you?

Go deeperBeyond checking work, verifiers pick winning answers and train models

In best-of-N, a model generates several answers and a verifier keeps the one it scores highest. Reinforcement Learning with Verifiable Rewards (RLVR) trains a model on tasks with checkable answers, such as math with known results or code with tests, and rewards it when the check passes.

Training makes any gap in a check a target. A model rewarded by weak tests can learn to pass them without solving the task, a form of reward hacking. Teams measure each check's false passes and false failures on cases with known answers, and compare full systems, since a strong model alone sometimes beats a weaker one with checks.

Practice Modern AI

Or take the learning quiz on a mix of every set