Retain upstream 32768-token input bound without truncation. CPU parity passed 12 cases, including exact 1024/2048 tokens, 1/2/3 options, and JSON. Standalone Python and native Swift verified. Native 32K inference remains untested; decision adapters remain outside this export. English and Chinese model cards describe the validated scope.
Use English model card by default and link Chinese translation
Fixed 128-token, two-option choice/noul FP32 graph. Six CPU source-parity cases passed; standalone Python runs without PyTorch; native Swift CPU inference verified. Includes Python/Swift examples, pinned uv environment, evaluation records and checksums. Other upstream adapters are not included.