Coding Before Testing Biases an LLM's Own Test Suite Toward Its Buggy Code

arXiv·medium signal

Finds that in agentic workflows where a single model writes code first and tests second, the resulting tests are biased toward passing the (possibly wrong) code, gutting their value as an independent oracle. It offers empirical support for enforcing test-first ordering in AI-assisted development. A concrete data point for a workflow decision builders make every day.