arXiv: scicode-lint — LLM-Generated Patterns Catch Methodology Bugs Invisible to Traditional Static Analysis

arXiv·low signal

scicode-lint targets a class of bugs conventional linters cannot detect: methodology errors in scientific Python that produce plausible but incorrect results — wrong statistical test, invalid normalization, violated assumptions. The tool uses LLMs to generate domain-specific lint rules from scientific literature and best practices, then applies them as executable checks against Python codebases. As coding agents increasingly generate domain-specific scientific and data-science code, this addresses a verification gap where syntax and type correctness provide no signal about numerical or statistical correctness.