Abstract
We describe GAIL-DDA, an interpretable controller for adjusting task difficulty in gamified skill-building activities. Rather than relying on accuracy alone, it folds a rolling accuracy measure, a completion-time adequacy term, and a webcam-derived observational expression signal into one weighted composite score, then decides whether that score is allowed to move the active difficulty level using a persistence-and-cooldown hysteresis rule. Because no human-subject data exists for the system yet, we validated it through a reproducible synthetic simulation built around four disclosed learner archetypes, five controller conditions, 40 matched random seeds, and 150 trials per condition-seed pair. Under the simulator and signal-generation assumptions documented in the accompanying reproducibility package, the full controller kept trials inside a predefined target-performance zone on 45.6% of trials, compared with 40.6% for an accuracy-only staircase baseline, and cut the mean absolute difficulty change per trial from 0.0889 to 0.0404. Both differences held up under matched-seed statistical testing. These figures differ somewhat from an earlier version of this manuscript: the original simulator code was not available to us, so it had to be rebuilt from the equations disclosed in the paper, and the numbers reported here come from that rebuild. What we report, then, is a simulation-level proof of concept and nothing more. We make no claims about human-subject usability, engagement, facial-expression accuracy, learning, or clinical or therapeutic outcomes in real children — human evaluation of the Feelio application remains future work.