Content Over Format: What Drives Analysis-Mode in Small Language Models, and How Grounding Recovers Execution

A prior report observed that a small language model, when a task is prefixed with a structured description of a recurring decision, stops acting and instead analyzes the description. That study used one model and one scenario, and it did not separate the description’s content from its form. This follow-up adds the controls it lacked. We prefix a task with either a three-axis descriptive “glyph” or an information-matched imperative rule, across four decision classes, and we sweep the descriptive prefix across four open-weight models on the original class. The shift away from action generalizes across models, and it is a content effect, not a format effect: the matched imperative suppresses action as much as the glyph, and the descriptive format shows no consistent advantage over a plain rule with the same content. Control prefixes (a content-scrambled glyph, a wrong-class glyph) show the suppression mechanism is decision-class-dependent. We then vary two things independently while holding the content fixed: form (descriptive versus directive) and grounding (a domain-free statement versus a domain-specific one). Grounding is the lever. A domain-specific aid restores execution where a domain-free one suppresses it, in both forms. Whether a domain-specific description suffices as well as a domain-specific command is under-determined here: the command reaches ceiling in all four model-and-class cells, so the two separate only where the description falls short of ceiling, which happens once, on the smaller model and the harder task, and there the command does better. All runs use one local raw-completion inference rig; reproducing them needs a 24 GB GPU and no paid interface. We state the scope and the measurement limits, including the cells where a baseline already acts and no effect is measurable.