[dev_tools_lab]

This is exactly why single-model evaluation is dangerous. Benchmarks are gamed, but disagreement between models is harder to fake. Multi-model consensus catches what individual benchmarks miss.