From Answer Production to Defensible Engineering Judgment in AI-Containing Systems: The iSCARB Framework and a Nine-Chapter Software Engineering Implementation

Abstract

Generative AI makes fluent artifact production inexpensive, weakening assessment validity when a submitted answer is treated as sufficient evidence of student capability. At the same time, software engineers increasingly build systems in which learned components sit inside larger socio-technical systems and can fail without conventional software failure signals. This preprint presents iSCARB, a course-level design that treats defensible engineering judgment, rather than answer production, as the observable assessment target. A defensible judgment requires the learner to select a fit option (FIT), state an observable boundary beyond which it ceases to be fit (BOUND), translate the judgment into action (ACT), provide inspectable evidence (EVIDENCE), commit to these before a changed condition (STRESS) is revealed, and then retain, revise or replace the original judgment with justification (REFIT). These mechanisms do not prove internal cognitive ownership; they create observable evidence that a learner can defend and revise a consequential decision. We report implementation across nine chapters of an undergraduate Software Engineering II course at King Abdulaziz University and an extension that threads an AI component through each chapter. The extension is intentionally a transfer layer rather than a new formal learning outcome: the five objectives, readings, self-checks and assessed assignments remain unchanged while existing software-engineering concepts are applied to AI-containing systems. Each chapter instantiates a different engineering stance toward the learned component, from doubting and measuring it to bounding, contracting, retrying and negotiating it across owners. We map the implementation to CS2023, report design iterations and automated quality assurance, and specify a planned evaluation of defensibility, revision, dependency reasoning, metric choice and instructional load. We make no claim of causal effectiveness, superiority or accreditation.