Verification that preserves the target model's output distribution and expected completion length by only accepting tokens the target would generate.