The Document Engineering textbook series (Vol 0– 5: Foundation, File Systems, Markdown, LaTeX, Citations, Diagrams), aimed at Vietnamese learners aged 14 and up, frames every pedagogical design choice on a single spectrum: at one end, augmentation (the AI integration leaves the learner stronger when the AI is removed); at the other, atrophy (the integration leaves the learner weaker). Vol 0 Ch. 2 establishes the spectrum conceptually; the series' content review process needs an operational rubric that classifies any chapter, section, or activity on that spectrum in finite time, without external evaluators, and at audit-loop scale. This paper specifies that rubric. Seven criteria are derived from the 2024–2026 empirical literature on LLMassisted learning and from the cognitive-architecture / desirabledifficulty / multimedia-learning anchor literatures: generativefriction preserved; verification step required; transfer-afterAI-removal tested; cognitive-load discipline observed; citation provenance explicit; reflective revision mandated; failure modes named. Each criterion is stated as a yes-no question scored at chapter, section, or activity scale; each carries a decision rule that maps the score onto the augmentation-atrophy spectrum (atrophy / mixed-risk / mixed-safe / augmentation); each is grounded in at least one empirical or theoretical anchor and at least one worked example from the Vol 0 draft. Calibration across reviewers is handled by yes-no item design and by anchor-example release rather than by inter-rater training. The rubric is consumed by the program-level audit loop (Gemini-as-judge, rubric-driven) as the per-artefact spectrum-position scorecard; it does not address external-evaluator workflows, learner self-assessment, or summative grading. Three open questions carry forward: weighting under inter-criterion trade-offs, transferability beyond the 14-and-up Vietnamese audience, and stability under postconversational AI interfaces.
That Le (Tue,) studied this question.