Use OpenPI-compatible normalization, token boundaries, balanced loss, and strict full-chunk decoding so training and inference share one sequence contract. Co-authored-by: Cursor <cursoragent@cursor.com>
pepijn