SafeCine: Risk-Calibrated Physical Preference Supervision for Language-Conditioned Aerial Cinematography

Language-conditioned aerial cinematography requires both expressive trajectory generation and physically reliable supervision. Near safety boundaries, independently perturbed rollouts can confound trajectory quality with disturbance variation, creating unstable pairwise labels. We introduce \emph{Risk-Calibrated Local Counterfactual Preference Mining} (RCL-CPM), which compares candidate trajectories generated from the same context under matched physical perturbations and abstains when finite-sample evidence cannot resolve the comparison. The resulting preferences align a diffusion proposal policy, while SafeCine keeps runtime execution decisions separate through physical and task admissibility checks before cinematic selection. On a frozen matched physical evaluation, the aligned policy improves lower-tail clearance by $0.11\,\mathrm{m}$ (95\% CI $[0.09, 0.15]$). Closed-loop evaluation reveals a gap between proposal quality and execution coverage, motivating a layered design in which preference supervision improves proposals while runtime verification governs execution.