A scoring mechanism that compares the probability of generated text under example-guided versus unconditioned models to favor demonstration-consistent outputs.