Use theorem provers to ensure the correctness of your LLM's reasoning (github.com)

A Python toolkit combines LLMs with the Z3 theorem prover to produce verifiable reasoning: the high-level ProofOfThought class wraps an OpenAI client to turn LLM-generated reasoning into Z3 constraints, query the solver, and return a logically-checked answer (example: a query returned False). An EvaluationPipeline lets you batch-evaluate on datasets (the example runs StrategyQA samples and reports an accuracy metric), and the project is structured as a two-layer system—a user-friendly API (z3dsl.reasoning) and a lower-level JSON-based DSL (z3dsl) that encodes Z3 problems. The repo includes examples (including Azure OpenAI support) and requires z3-solver, openai, scikit-learn, and numpy. For AI/ML practitioners this hybrid symbolic–neural approach is significant because it adds formal verification to chain-of-thought style outputs, enabling answers to be checked, counterexamples to be found, and overall trustworthiness to be measured rather than only plausibility judged. Technically, the system translates LLM reasoning into SMT constraints that Z3 can prove or refute, making it suitable for complex logical QA, safety-sensitive applications, and benchmarking reasoning quality. The toolkit supports evaluative pipelines and reproducible experiments, so teams can both generate and certify reasoning at scale.