Modeling guilty minds: Laypeople treat knowledge as categorically distinct from recklessness

The Model Penal Code (MPC) distinguishes four levels of criminal culpability (purpose, knowledge, recklessness, and negligence), yet whether laypeople can reliably make these distinctions through genuine inference remains unclear. Prior studies either presented participants with explicitly labeled mental states or asked for self-assessed judgments that may not reflect natural reasoning. We addressed this gap by requiring 212 participants to infer a protagonist's mental state from behavioral evidence in contraband smuggling vignettes with six systematically varied binary features, including the type of contraband and whether payment was offered. Participants classified the protagonist as blameless, negligent, reckless, or knowing and rated the perceived risk of the situation. Cumulative link mixed models showed that risk assessments strongly predicted culpability categorizations and that payment was the only vignette feature to independently shift judgments toward higher culpability after correction for multiple comparisons. Critically, knowledge judgments were associated with lower risk ratings than recklessness judgments, and the proportional odds assumption was violated, with the risk coefficient at the knowledge threshold substantially weaker than at lower thresholds. A features-only Random Forest with SHAP analysis revealed that knowledge was the only category where the type of contraband rather than payment dominated feature importance. Representational similarity analysis confirmed that semantic distance between vignette texts predicted judgment distance, even though the features driving semantic similarity differed from those driving judgments. These results indicate that laypeople can systematically distinguish MPC mental states through inferential reasoning, but that knowledge functions as a qualitatively distinct category driven by categorical inference rather than extreme risk perception.