Munim, Inc. developed "eval-reliability-audits," a catalog of 25 reproducible audits designed to identify defects in LLM benchmark scoring code and gold labels. Using sealed containers and pre-registered synthetic fixtures, the researchers isolated specific errors to measure their impact on released data. In Natural Language Inference datasets, they identified conflicting gold labels for identical premise-hypothesis pairs: SNLI contained 64 exact and 82 normalized conflicts, while MultiNLI had 83 exact and 86 normalized conflicts, all within training sets. An audit of GSM8K revealed that while calculator annotations are 100% accurate, the final #### answer deviates from the last computation tag in 5.4% of solutions (470 of 8,679). Analysis of GSM8K's "socratic" version showed 99.93% formatting consistency. Furthermore, the study uncovered flaws in scoring logic: HuggingFace's exact_match normalization can flip verdicts based on punctuation, and the ROUGE tokenizer's ASCII filter may yield a ROUGE-1 F-score of 1.0 for summaries with opposite meanings. IFEval's checkers exhibited latent hazards, including missing Unicode equivalents and regex failures in keyword checkers, though these often had a zero blast radius on the original 541 prompts.