From Training to Qualification: A Framework for AI Education and Competency Assessment

An academic approach to artificial intelligence (AI) education requires more than organizing training material into courses or assigning certificates to benchmark scores. It requires separate evidence that an intervention produces learning, that an identified system meets a competency standard, and that the resulting qualification supports its intended use. This paper develops a framework that connects these three claims without treating them as interchangeable. Drawing on curriculum learning, evidence-centered assessment design, argument-based validation, and AI-system documentation, the framework defines the assessed student as a version-bound model configuration rather than a model brand. It separates educational records from competency transcripts, distinguishes parameter-based learning from reference-assisted performance, and specifies independent examination, bounded qualification claims, and reassessment after material changes. A prospective pilot, ER-101: Evidence-Grounded Research Assistance, operationalizes the framework using two open-weight model families and four conditions: an unchanged starting configuration, handbook assistance, block-shuffled course training, and prerequisite-ordered course training. The trained conditions use identical optimizer-step bundles, example exposures, and update budgets; only first-pass bundle order differs. Three paired training seeds per family yield twelve trained checkpoints and sixteen evaluated configurations. A protected 300-case examination is separated from 60 practical assignments that cannot influence qualification decisions. The planned analysis distinguishes task-sampling uncertainty from training-seed variability and treats qualification-to-practice correspondence as exploratory. Analytical planning calculations illustrate why moderate task counts cannot establish rare-error safety or population-wide credential validity. No pilot data, trained checkpoints, qualification outcomes, or empirical effectiveness results are reported. The contribution is a falsifiable education-to-qualification architecture and a controlled protocol for investigating it, not a claim of human-equivalent degrees, universal trainability, or validated professional competence.