38% of Content-Correct LLM Markdown Outputs Have Broken Code-Fence Boundaries

arXiv 2609.06993·medium signalXBlueskyLinkedInCopy link

LatentMD separates content correctness from boundary correctness in CommonMark fence handling, using 4,179 prompts and a CLI for scoring arbitrary model outputs. Across 9 LLMs and roughly 37,600 generations, 38.0% of valid main-grid outputs are content-correct but boundary-broken, with substantial breakage under unspecified prompts and in a human-authored validation set. Ablations attribute failures primarily to same-family symmetric-delimiter collisions rather than nesting depth, show prompt hints only partially mitigate them, and confirm the problem generalizes to Python triple-quote docstrings while JSON stays robust as an asymmetric-delimiter control.