Error-injection exercises: diagnosing subtly wrong simulations as graded practice¶

Strategy · All strategies
Evidence · no claims cited

Description¶

The article recommends giving students a subtly wrong simulation—sign error in the force, dt an order too large, the wrong potential—with the diagnosis as the graded object. The author states the exercise "provides the supervisory practice that the agentic workflow no longer provides by itself," and trains reading code one did not write, the fluency the oral defense's walkthrough demands. Several of the author's in-class quizzes have taken this form for years, and the richest source of injected errors is now real agent sessions. Its complement is computing education's prompt problems, where students author the specification.

Design Implications¶

Context¶

Requirements¶

• The exercise presupposes the verification repertoire of Sec. IV, since diagnosis requires knowing what counts as evidence of correctness.

Constraints¶

• The article does not report systematic outcome data for this exercise; it is offered as practitioner practice within an anecdotal account.

Target Learners¶

• undergraduate physics students in computational courses

Target Learning Goals¶

• reading and debugging code one did not write; diagnosing simulation errors

Related Strategies¶

• Erroneous Examples
• Use synthetic LLM-generated error datasets for AI tutors, teachable agents, and teacher professional development
• Emphasize Reasoning over Answers
• White-box-then-black-box sequencing: transparent study of a method before opaque delegation

Examples¶

-

Key Sources¶

• Larry Engelhardt. (2026). A Tool-Invariant Framework for Teaching and Assessing Computational Methods in the Age of Agentic AI. https://arxiv.org/abs/2607.15518