Researchers reported a dual-pathway deep learning system that predicts tuberculosis drug resistance with over 95% accuracy while auditing sequencing quality and recovering both canonical and novel resistance loci. The model was built to reduce the risk of performance drop-offs caused by data artifacts. The emphasis on self-auditing addresses a persistent issue in genomic AI tools—models can perform well on curated datasets yet degrade when sequencing inputs vary in quality. Here, the model’s design includes mechanisms to check evidence before producing resistance calls. If validated prospectively, such systems could accelerate resistance-aware treatment planning, particularly where rapid phenotyping capacity is constrained.