Instruction-tuned large language models are vulnerable to collapse when faced with trivial constraints, such as the removal of a single punctuation character or common word. This fragility is evident in the significant loss of comprehensiveness in their responses, with losses ranging from 14% to 48% across three open-weight model families. The collapse of helpfulness is observed even when the constraints are minimal, highlighting the brittle nature of these models. Researchers have demonstrated this phenomenon through pairwise evaluation, showcasing the sensitivity of instruction-tuned LLMs to simple lexical constraints1. The implications of this finding are significant, as it suggests that these models may not be reliable in real-world applications where unexpected constraints may arise. This vulnerability matters to practitioners, as it underscores the need for more robust and resilient language models that can maintain their helpfulness even in the face of minor constraints.