Balan, A.; Elhanati, Y.; Meza Landeros, K. E.; Mendes, M. D. A.; Lai, J.; Zaidi, S. S. A.; Unal, M.; Kim, B. Y. S.; Lucas, C.-H. G.; Runco, E.; Puduvalli, V. K.; Gantchev, J.; Whittaker, C. A.; Sharma, P.; Tabar, V.; Cima, M. J.; Baquer, G.; Reardon, D. A.; Stortchevoi, A.; Boire, A.; Wang, L.; White, F. M.; Sidiropoulos, D. N.; Yu, K. K. H.; Chiocca, E. A.; Anagnostou, V.; Data Science Teamlab, ; Accelerating GBM Therapies TeamLab, ; Karchin, R.

T-cell receptor (TCR) repertoires encode the organization of adaptive immunity and its reshaping by cancer and therapy, but disentangling treatment-associated structure from V(D)J recombination constraints remains challenging. We present CRAFT (Cancer Repertoire Anomaly Finding Transformer), a conditional sequence-to-sequence transformer that learns a nucleotide-level generative grammar of productive TCR-beta CDR3 sequences from healthy-donor repertoires, conditioned on germline V(D)J assignments. A dual-head decoder mirrors the independence of V-D and D-J recombination, and curriculum training yields embeddings that serve as a reference coordinate system for quantifying structured deviations in cancer-associated repertoires. In proof-of-concept analyses of a checkpoint blockade cohort (n=18) and a two-patient single-cell study of oncolytic immunotherapy, CRAFT-derived geometric metrics capture response-associated immune remodeling, including longitudinal shifts in repertoire organization. In antigen-labeled benchmarks, CRAFT yields coherent organization across specificity classes while highlighting settings where CDR3-beta alone provides partial signal.