Laisan86

ExpireMem Guardian — Verify Before You Execute AI coding agents can generate patches that appear correct but fail because an API has been removed. More dangerously, a patch may execute successfully while silently changing or corrupting data. ExpireMem Guardian is a version-aware safety and verification workflow that evaluates AI-proposed code repairs before accepting them. It combines API lifecycle rules, restricted source-code inspection, official documentation evidence, and execution-based tests. When compatibility or safety cannot be established, the system abstains rather than blindly trusting a proposed fix. In a controlled pandas example, code using the removed "DataFrame.append()" API fails under pandas 2.2.3, while a constrained "pd.concat()" replacement passes the fixed test. Additional gapped-index tests demonstrate why a successful exit code alone cannot guarantee data correctness. The project includes a six-tab local Streamlit Studio, a curated 14-case version-rule benchmark, reproducible tests, and auditable evidence. A historically recorded Meta Muse Spark repair proposal through Baseten is distinguished from the public offline replay. The demonstration does not require new model calls or paid inference. Our goal is to make AI-generated code repairs more verifiable, not merely executable.