Constraint-Structured Prompt Grammars and Their Effect on Transformer Reasoning Signatures

This study investigates whether constraint-structured prompt grammars affect the structural distribution of reasoning features in transformer outputs. Using a controlled experiment with Claude Sonnet 4, eight reasoning features were measured across constraint-structured and unconstrained prompt conditions on a decision governance task. Seven of eight reasoning features differed significantly across grammar conditions (Bonferroni-corrected alpha = 0.00625). The largest effect occurred in structural reasoning (Cohen's d = 4.58, 95% CI 3.8, 5.4). Total reasoning volume showed no statistically detectable difference between conditions (ANOVA p = 0.069), suggesting prompt grammar redistributes reasoning features rather than increasing reasoning quantity. These findings support the hypothesis that prompt grammar influences the structural organisation of reasoning in transformer outputs and may function as a control surface shaping reasoning signatures.