Derived Standards for LLM-Assisted Engineering Student Workflows

Engineering students increasingly use generative AI tools built on large language models (LLMs) for coding, calculations, literature searches, and technical writing. Yet university policies often emphasize whether these tools may be used and how their use should be disclosed, offering less guidance on how students should verify their outputs before relying on them. We propose a practical verification standard to address this gap, developed by a team that includes engineering students who use LLMs in their own work. Drawing on a narrative review of published research, we make four contributions. First, we identify where errors can enter student work across five common task categories. Second, we propose a taxonomy of eight error types, ranging from misinterpreted instructions to citations that do not support their associated claims. Third, we propose three requirements: record what was asked, verify outputs before relying on them, and state what was and was not checked. A central safeguard requires at least one verification step that does not depend on an LLM, because evidence indicates that LLMs may fail to detect their own mistakes and that different models tend to make the same mistakes. Fourth, we translate these requirements into STAPLE (Save, Trace, Anticipate, Protect, Limit, Enable), a six-step checklist that scales verification effort according to the consequences of error. The proposed standard has not yet been empirically evaluated, and we outline studies needed to assess its usability and effectiveness. Our aim is to establish verification practices that transfer from engineering education to professional practice, where unchecked LLM errors can have consequences far beyond academic performance.