Two-instrument assessment: AI-free coding quizzes plus verification-gated oral defenses of comment-stripped AI-assisted work¶

Strategy · All strategies
Evidence · no claims cited

Description¶

The article's practical response to proxy collapse pairs weekly AI-free coding quizzes, which assess the white-box phase, with ten-minute one-on-one oral defenses of AI-assisted, comment-stripped project work, which assess orchestration. The author argues "if coding fluency is a learning goal, AI-free conditions are now the only conditions under which it can be validly measured," and the defense is scored on a five-dimension rubric with a gate: the verification dimension must reach the functional level to pass. Live adaptive probing, not comment-stripping, is what certifies, since the probing does not exist in advance and cannot be rehearsed.

Design Implications¶

Context¶

Requirements¶

• Requires small class sizes: ten minutes per student per defended assignment, roughly twenty-eight contact-hours per semester at the projected fifteen-student cohort; degraded modes (spot-defenses, TA-led defenses, paired defenses) each carry named costs.

Constraints¶

• The comment-stripping, published rubric with verification gate, and sophomore-level population are genuinely untested until Fall 2026; the rubric thresholds are one examiner's unvalidated defaults, and the format does not scale to a three-hundred-seat course, raising an equity concern.

Target Learners¶

• undergraduate physics students, sophomore-level computational course and upper-level courses

Target Learning Goals¶

• explaining and defending computational artifacts in the language of the discipline; calibrated reliance on AI

Affordances¶

• Five Pillar Tool Invariant Computational Framework

Related Strategies¶

• Pair AI-permitted take-home assignments with supervised modifications, oral defense, prediction, and transfer tasks to verify understanding
• Use weekly oral code review interviews to incentivize students to understand AI-assisted code
• Mini homework defenses: short oral one-on-one assessments of student solutions
• Replace detector-centred assessment with a four-stage process-oriented model: declared starting point, versioned drafting, process artefacts, and targeted oral verification
• Redesign assessment toward supervised, oral, process-based and authentic formats to verify students' own reasoning under GenAI

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