In order to ultimately build an automatic evaluation system to predict voter errors, cognitive models have been developed to simulate a family of possible voters’ behaviors. However, the previous models cannot interact with non-race content (e.g., ballot instructions, header, footer) on full-face paper ballots, while the interactions of voters with non-race content can sometimes lead to voter errors. For example, on the ballot used in Broward County, Florida in 2018, a long instruction section and two races were placed in the left column. Some voters regarded the left column as an instruction column and skipped the left column entirely. This ballot design caused more than 26,000 people to miss the Senate race listed right below the instructions. The purpose of this research is to expand the strategy space and to update the previous models to interact with non-race content. This research consists of two eye-tracking experiments and a model. The primary findings of the experiments are as follows: the design of voting slates influenced the reading patterns of participants; most participants spent limited time on ballot instructions; the length of ballot instructions did not affect how much time participants spent on instructions. The insights from the eye-tracking experiments were then integrated into a new set of voting models. A total of 360 models were developed and the Broward error was reproduced successfully.