Connectomes provide critical constraints for neural circuit models. Advances in electron microscopy have dramatically increased the scale of connectome datasets, enabling more realistic modeling of neural circuits. However, challenges remain in using these datasets. Many models still rely on unconstrained physiological parameters, even within fully mapped circuits, leading to ongoing debates about the predictive power of connectomic data. Key questions remain on how synaptic connectivity relates to physiological properties and the anatomical detail needed to predict function. One modeling choice is to constrain connection strength using the absolute number of synaptic inputs. A second choice is the relative number of synapses---a fraction of total inputs from specific partners. Finally, a third choice is to use synapse size, which correlates with the magnitude of the postsynaptic potential. Distinguishing which parameter, or combinations of parameters, improves the predictive accuracy of connectomic-constrained models is difficult because it requires wiring diagrams and synapse size measurements between uniquely identified partners. To address these challenges, we developed a rate-based model of a fully mapped and tractable motor circuit in the ventral nerve cord of Drosophila larvae, incorporating known behavioral data. Initially constrained by connectivity and absolute synapse count, we incorporated relative synaptic weights into our models and optimized their scaling to predict circuit function. The model with relative weights outperformed those based solely on synapse count, accurately assigning functional roles to individual neurons. Additionally, we quantified postsynaptic density areas to correlate the scaled weights with synapse size observed in real data, validating their biological relevance. Finally, we explored connectomic constraints using mutants with shifted neuron locations or blocked synaptic transmission to further assess the predictive power of such parameters in perturbed circuits. These findings demonstrate that incorporating anatomical factors like synapse size significantly enhances the predictive power of neural circuit models.