Supplementary Material and Reproducibility Package: The Handover Problem — Governing Autonomy Transitions in Human–AI Collaboration

This deposit accompanies the paper "The Handover Problem: Governing Autonomy Transitions in Human–AI Collaboration" (under review, IEEE Transactions on Human–Machine Systems). It is a reproducibility package containing the full supplementary document, all aggregated simulation outputs, and the scripts that generate every figure and secondary analysis in the paper.

The paper formalises the Handover Problem — the challenge of governing when and how control transfers between human and AI agents across multi-cycle collaborations — and proposes the Handover Readiness Score (HRS), a transparent composite index integrating four signal dimensions: human readiness, relational trust, learning stability, and operational performance. A hysteresis-based transition policy uses the HRS to govern autonomy escalation, maintenance, and reversion, complemented by hard safety guards. The central finding is complementarity: guards catch abrupt single-signal threshold breaches, while the HRS detects the slow, multi-signal erosion of operator readiness that no individual guard can observe.

The deposit includes:

• handover_problem_thms_supplement.pdf — the supplementary document in five sections: S1 component ablation, S2 shock-severity sweeps, S3 learning- and operational-stability shocks, S4 threshold and weight sensitivity analyses, and S5 the healthcare cross-domain boundary analysis.
• data/ — 13 CSV files with all raw simulation outputs: steady-state performance and sustainability metrics (H1, H2); gradual-deterioration detection coverage, latency, and comparators (H3); HRS trajectories and component breakdowns; threshold and weight sensitivity grids (H4); the silent-AI-degradation (overtrust) experiment; and mode-change frequency.
• scripts/ — 8 Python scripts that reproduce all figures and run the secondary experiments.
• CODEBOOK.md — variable definitions and file schemas for every CSV.

Experiments use three conditions (BASELINE: LinUCB allocator, no governance; GOVERNANCE: cap/guard layer; READINESS: cap/guard + HRS) across software-engineering and manufacturing domains (50 seeds, T = 100 cycles), with a healthcare boundary analysis. The primary benchmark studies gradual sub-threshold skill erosion; a supplementary stress-test family applies direct signal-injection shocks. All data were produced on HAAS Studio, a Human–AI Adaptive Symbiosis research platform; a live instance is accessible at http://3.248.226.75:8501. HAAS Studio source code is available from the corresponding author on reasonable request.