PureCheck – A Blank-Slate Inference-Step Verifier

PureCheck is a new type of inference-step verifier designed to help AI models identify gaps in their reasoning. After a large language model generates a chain of thought, PureCheck checks each transition to flag any step where the conclusion does not logically follow from the premise. It acts as a post-hoc logic auditor that highlights unsupported leaps with a score and an optional hint. Unlike other AI checkers, PureCheck resets its memory to zero before and after each check, ensuring a truly “blank slate” with no retained context or facts. It trains itself through self-play, creating countless tiny reasoning puzzles and learning to catch hidden logical gaps: it earns rewards for flagging real flaws and penalties for false alarms. Over time, PureCheck becomes finely tuned to spot only genuine reasoning leaps. Key benefits of PureCheck include: Accurate logical checks: It flags true reasoning leaps with high precision and very few false alarms. Lightweight and fast: The model is small (around 100 million parameters) and efficient, running on a single GPU at dozens of chains per second with minimal latency. Easy integration: PureCheck is provided as a plug-and-play API and library. Other AI systems can send their chain of thought for verification and quickly receive flags with coherence scores and hints. Zero retained bias: By wiping its state after every run, each check starts fresh. This memory reset ensures the verifier’s judgments are unbiased by any prior queries. PureCheck’s narrow focus on step-by-step logic provides a powerful guardrail for complex AI systems. It catches subtle reasoning slips that larger models might overlook, improving overall trust in AI outputs. Due to its focused task and small size, PureCheck can be open-sourced as a compact, transparent module that any developer can attach to their AI workflow for rigorous reasoning checks. Note: This whitepaper presents a conceptual overview of PureCheck; it does not include experimental results.