Lim

A passing test is not the same as a correct test. When an AI writes tests from buggy code, the tests can pass while locking the bug in. Autonomous QE Agent generates pytest tests with Meta's Muse Glimmer, then writes an independent reference implementation from the function's written spec (its docstring) without seeing the code body. It compares the source, the tests and the reference in plain code to decide whether "the test is wrong" or "the code is wrong", audits tests that pass, and exports a spec-aligned test suite that fails until the bug is fixed. Every model call goes to Muse Glimmer 30B (open-weight, 4-bit, llama.cpp on a rented GPU) and is shown live with the model name and timing. On my own benchmark of injected bugs it found 21/21 on the main set and 9/12 on a held-out set, with 0 false alarms on 12 clean versions (one run each; the bugs are simple and self-made). Text outputs are a known blind spot.