Governed Human-AI Prioritization Under Uncertainty - Aggregate Benchmark Results for the Revised Manuscript

This repository provides the public aggregate benchmark results accompanying the research manuscript "Governed Human-AI Prioritization Under Uncertainty: Reliability-Aware Estimation and Dependency-Constrained Portfolio Selection" by Azzeddine Ihsine and Sara Ihsine (Inovionix). The study investigates the reliability, robustness, and limitations of quantitative decision operators used within D-POAF, a governed framework for AI-native software engineering. The research evaluates five interconnected areas: 1. Governed prioritization: Sensitivity of Business Value Scores (BVS), Effort and Risk Scores (ERS), and Prioritization Value Scores (PVS) to changes in scoring parameters and their effects on prioritization decisions. 2. Effort estimation and forecast combination: Comparative evaluation of individual estimators, inverse-MSE weighting, covariance-aware Bates-Granger combination, regression-based aggregation, and shrinkage methods under correlated estimation errors. 3. Human-AI calibration: Evaluation of model-collective divergence as a diagnostic mechanism, including situations where human and model-based estimates share outdated contextual information. 4. Dependency-constrained portfolio selection: Comparison of transparent selection heuristics against exact Mixed-Integer Linear Programming (MILP) solutions under effort budgets, risk-adjusted objectives, and prerequisite constraints. 5. Optimal Development Path (ODP): Evaluation of the native dependency-aware sequencing rule against an exact mathematical reference on controlled synthetic instances. The experimental evaluation combines reproducible simulations using 30 deterministic seeds with analyses of four public software-effort datasets: Kitchenham, Desharnais, China, and Maxwell. The historical datasets provide empirical evidence concerning estimation accuracy and error correlations. They do not contain AI-generated estimates or represent completed D-POAF Waves. Portfolio-selection and sequencing experiments are evaluated separately using synthetic instances. The public archive contains aggregate experimental results, comparative tables, figures, and documentation describing data provenance, methodological assumptions, and access conditions. This repository does not contain proprietary D-POAF platform source code, full experimental implementations, or confidential project information. Additional experimental materials, including detailed configurations and research verification scripts, may be requested from the corresponding author, Sara Ihsine ([e-mail]), subject to authorization, applicable licensing conditions, and data-access restrictions. The results characterize the conditions under which governed prioritization, reliability-aware estimation, and dependency-constrained selection methods perform effectively, as well as situations in which they may fail. They provide quantitative evidence supporting transparent and auditable software-planning decisions. They do not establish causal improvements in actual project costs, calendar delivery times, or realized business value.