Abstract
Hybrid artificial intelligence optimization models combine learned behavioral signals with constrained decision-making, but their transition from prototype calibration to operational use is rarely evaluated across changing platform regimes. The Adaptive Calibration and Operational Transferability (ACOT) framework separates the parameterized decision utility from a fixed external evaluation instrument, validates calibrated configurations on unseen stochastic scenarios, and audits directional transfer regret. The framework is evaluated on a de-identified pilot dataset from a digital group-buying platform comprising 150 users, 200 products, 150 lots, 4000 behavioral events, and 500 orders. Because the observed lot records predominantly represent completed or expired states, the assignment experiments use counterfactually reconstructed pre-activation lot states rather than a complete historical replay. A corrected objective-aware heuristic restores identifiability of the relevance–completion parameter. Across 20 matched optimizer seeds, Tree-structured Parzen Estimation achieved mean validation J=0.734518, compared with 0.733899 for random search and 0.730085 for expert weights, and won 15 of 20 seed-paired comparisons. The run-level BCa 95% confidence interval for the mean TPE–random difference was [−0.000001, 0.001200], indicating a modest, seed-sensitive advantage. Continuous stress testing showed a non-monotonic calibration value, with the largest sampled gain occurring at an available-user fraction of 0.90. Tight-budget calibration had the lowest point-estimate worst-case transfer regret (0.000720). However, hierarchical bootstrap assigned it a 55.1% probability of being the minimax source, compared with 40.1% for the base regime. The results support uncertainty-aware transfer auditing rather than assuming a universally robust calibration regime.