sayanpal1499/VerityBench ? reverse-engineered prompt

Build me a Python project that evaluates how reliable an LLM is at answering questions from documents.

I want it to take a small set of source documents, generate a mix of normal and adversarial QA cases, run the model, and score things like factual correctness, when it should abstain, and whether it hallucinates unsupported claims. It should keep the whole process auditable, with clear saved outputs, human review markers for questionable cases, and reproducible reports in the repo.

Please make it work without relying on big evaluation frameworks, and include a solid offline test suite so I can verify the logic without any API calls. If you need current docs or SDK details, look them up online. I also want a strict verification step so live benchmark runs stop if the dataset records are not approved yet.

Use plain Python, save results and reports locally, and make the output easy to inspect later.