October 13, 2026 local evaluation: raising the constrained generation cap from 96 to 160 tokens corrected the remaining truncated output, improving structural validity from 209/210 to 210/210 on the original seven-task regression set. All other 209 outputs were unchanged. A separate 14-prompt paired check passed at both caps with identical output text. The original baseline separately recorded 209/210 constrained versus 97/210 retry outputs at the 96-token cap; the retry arm was not rerun at 160 tokens. These checks measure structural validity within the supported schema subset, not factual answer accuracy.

Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.