Disentangling Retention and Transfer in Resource-Aware Clinical AI Simulation

Clinical AI education should develop judgement that persists and transfers across systems, yet changing the AI system only at follow-up confounds transfer with forgetting. We developed a crossed assessment framework and evaluated its statistical properties using 5,000 Monte Carlo replications per scenario. Sixty synthetic learners per replication completed four system-by-time assessment cells, with 15% independent attrition, 20 decisions per cell and a shared learner effect. Familiar-system accuracy declined from 0.75 to 0.67; unfamiliar-system penalties ranged from zero to ten percentage points. When the penalty was five points, the diagonal comparison misattributed an additional 4.98 points to forgetting and achieved only 30.42% interval coverage of retention. Within-system retention estimates maintained approximately 95% coverage. A deterministic example showed that improved override alone can conceal complete rejection of safe advice. An accompanying resource audit corrected illustrative expenditure calculations while identifying unsupported energy comparisons. These computational findings support identifiable assessment contrasts and separate acceptance and override outcomes. They do not demonstrate adaptive-training effectiveness, clinical benefit or measured environmental savings; those claims require prospective validation.