delivery using a continuous synchronized beam delivery system. Continuous tracking of three LED markers, in conjunction with the timedependent radiographic position of the fiducial, was used to build correlation models to guide robot's compensation of the first-order target motion. Internal data from the system log-file was accessed to obtain marker model point positions, radiographic target positions, and the positional correlation error at the series of periodic time-points, analysis of which yielded a measure of the tracking accuracy. Results: Comparison of the 178 stereoscopic images defining the true target position with the 496 model points used for tracking yielded a mean positional correlation error of 1.8mm, 1.1mm, 1.8mm in x, y, z (system coordinates), with respective standard deviations of 1.4mm, 0.9mm, and 1.4mm. The mean radial 3D correlation error for the entirety of the treatment was 3.2mm with a standard deviation of 1.6mm; 90% of points had less than 5.5mm radial deviation, with outliers constituting model rebuilding. Conclusion: While phantom and animal experiments have produced preliminary data on the cardiac tracking under model conditions, the study presents the first in-patient determination of tracking accuracy for VT radio-ablation. Considering the mean magnitude and standard deviation of the 3D positional correlation error, a minimum 5mm margin should be used to account for accuracy of such surrogate-based tracking delivery.