InclusionAI's SingProbe adds risk scoring inside the generation process of the Qwen model Qwen3.8-27B. A small probe reads the model's internal calculations as it produces each tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →, scoring request intent, unsafe output and hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition → risk. The model card describes this as a way to monitor an unfolding response rather than running a separate safety model.
Deployment requires the exact matching base model and one of the project's serving integrations. Those integrations return a score dictionary for each generated token. The concrete capability is continuous risk assessment during generation; the scores are predictions, rather than verification that an answer is true or safe.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
