Built a little demo of using Jev to check against inefficient code evolution using standard agentic LLMs.

Built a little demo of using Jev to check against inefficient code evolution using standard agentic LLMs. How it works: a task involving programming happens as usual through codex, but edit tool_calls are intercepted, the code [edits] are statically analysed using fbinfer to obtain a work estimate differential and then a schema is prepared with instructions for Jev to classify at a fine-grained level the edits using the user-prompt and the code comments as well as the Infer runtime instrumentation output to score with respect to computational complexity, necessity for the requested change and accrual of poor architecture (the main failure antipattern is that as users request new features they are near-sightedly implemented in a way that makes the whole code inefficient because it doesn't trigger reappraisal of the architecture). If the classification is below some thresholds then an intervention is made with the Jev feedback and the