Boundary Recognition Through Lightweight Projects: Cultivating Engineering Judgment in a Compressed Numerical Methods Course Amid AI Tool Proliferation

ABSTRACT

Engineering education faces a dual crisis: foundational courses undergo severe hour compression while AI tools proliferate, creating tension between technical implementation and professional judgment development. This study proposes a lightweight project framework that addresses this dilemma within 24 instructional periods (45 min each; 18 contact hours in total) by strategically leveraging professional authenticity to cultivate boundary recognition —the ability to diagnose when computational tools exceed valid operating ranges. Guided by outcome‐based education and informed by situated learning perspectives, the framework operationalizes three design principles: (1) technical minimalism (core code < 100 lines), (2) professional anchoring (algorithms embedded in safety‐critical decisions), and (3) dual‐model contrast (structured comparison between physics‐based and data‐driven approaches). A quasi‐experimental study involving 263 civil engineering undergraduates showed substantially higher post‐reform scores across all assessed dimensions using Welch's independent‐samples t ‐tests (all p  < 0.001). The largest differences were observed for engineering risk awareness (mean difference = 2.66, d  = 4.83), uncertainty quantification (2.38, d  = 3.61), and critical AI usage (2.58, d  = 4.35). Sensitivity analyses adjusting for prior GPA yielded virtually unchanged estimates. Crucially, 81% of students established explicit decision rules linking sample size to appropriate modeling strategies, consistent with the professional competence principle articulated in the ASCE Code of Ethics. Thematic analysis of eight student interviews indicated that small‐sample constraints functioned as cognitive scaffolds by limiting non‐essential implementation demands and foregrounding uncertainty interpretation as an engineering responsibility. This work contributes a theoretically grounded mechanism model. It explains how professional contexts transform abstract algorithmic properties into tangible engineering boundaries. By doing so, the study moves beyond descriptive accounts of “what works” to explain how the approach may support boundary recognition in AI‐augmented engineering education.