Biomechanical models of eye kinematics are essential tools for decoding complex ocular behavior. These models can provide a theoretical framework for clinical or augmented-reality applications, such as predicting surgical outcomes in strabismus by simulating the mechanical realignment of extraocular muscles, or measuring visual fatigue in virtual-reality settings [2]. Traditionally, these models employ 1-D Hill-type muscle formulations, which offer computational efficiency but rely on highly simplified muscle geometries and idealized insertion points. While these models have been used in inverse-dynamic frameworks to derive the neural commands underlying saccadic movements [2], the impact of geometric simplifications on such commands remains poorly understood, and neural command estimates are often physiologically invalid. Given that neural commands for saccadic movements are prototypical [3], this work investigates the sensitivity of the estimated neural commands to variations in muscle geometry. This is done by implementing an inverse neuromechanical model of eye kinematics within the SOFA framework, which is then used to compare the neural command required for horizontal saccades across varying geometric configurations. The models were evaluated using a large-cohort of horizontal saccades from the open-access eye-tracking dataset GazeBase [4]. Our findings indicate that neural excitation features (e.g. pulse height, lead time, and the pulse-step transition) are significantly altered across the evaluated muscle geometries. These results demonstrate that geometrical simplifications must be handled carefully. Since geometric errors propagate directly into the estimation of neural commands, high-fidelity modeling of muscle paths is a prerequisite for developing reliable, patient-specific simulations. Future work will integrate these geometric insights to improve the predictive accuracy of surgical planning for strabismus and other oculomotor pathologies.