Repeatable output is not verified evidence

A local Qwen3 study repeated every model outcome and validated every citation. All 18 runs still failed admission. A frozen claim-skeleton compiler narrowed the model's job, yet 12 of 18 new runs remained refused.