The zero-to-one likelihood reflects the model's sigmoid or softmax probability that a specific word token functions naturally as anchor text, allowing the client-side confidence slider to group high-scoring adjacent spans without repeated network requests.