A verifier is a deterministic or objective computational system that checks whether an output satisfies a correctness criterion. It returns a passing or failing signal, often 1 or 0, that can measure an answer or serve as a reinforcement-learning reward. section 7.1.1 section 7.7.3
Use a verifier when the task has a correct outcome, correctness can be checked programmatically, and the criteria are objective and unambiguous. If any condition fails—as with tone, helpfulness, or open-ended faithfulness—a fixed verifier score can reward shortcuts rather than task success. section 7.1.1 section 7.7.3
Design a verifier by matching the check to the output:
- Compare a mathematical final answer with ground truth. section 7.7.3
- Execute generated code against a unit-test suite. section 7.7.3
- Validate structured output with a regular expression or schema-like format check. section 7.7.3 section 7.1.1
- Reserve preference-based signals for criteria that lack complete programmatic ground truth. section 7.7.3
Financial research assistant
A verifier can check whether a response cites the source, matches the required JSON structure, and reproduces a revenue figure from the retrieved transcript. These checks convert otherwise manual scoring into explicit reward signals. section 7.7.3
Table 7.1 Task types mapped to reward types
|
Chapter 1 task type
|
Chapter 7 reward type
|
|
Verifiable task
|
Programmatic verifier
|
|
Open-ended task
|
Preference-trained reward model
|
Task type determines reward type: verifiable tasks use programmatic verifiers, while open-ended tasks use preference-trained reward models. section 7.1.1








