mathura-snth/math-llm-verifier

The active product is a reproducible benchmark generator for evaluating whether language models detect logical errors in theorem hypotheses. A user runs the benchmark, which selects theorems and error types, generates and formally validates synthetic copies, and writes labeled cases to JSON. The formal model comprises theorem requirements, hypothesis text, and implication rules. The repository also preserves an earlier French-language implementation and a Tableau exporter. The main workflow is evidenced by sampled active source; the archive is shown as a separate historical branch, not part of the active pipeline.