Jason D. Yeatman ^1* Jasmine E Tran ² Amy K Burkhardt ¹ Wanjing Anya Ma ¹ Jamie Mitchell ¹ Maya Yablonski ¹ Liesbeth Gijbels ³ Carrie Townley-Flores ¹ Adam Richie-Halford ¹

• ¹ Stanford University, Stanford, United States
• ² University of California, Irvine, Irvine, California, United States
• ³ University of Washington, Seattle, Washington, United States

The speed at which students can accurately read and understand connected text is at the foundation of reading development. Timed reading measures go under a variety of names (e.g., reading fluency, reading efficiency, etc) and involve different levels of demands on comprehension, making it hard to interpret the extent to which scores reflect differences in reading efficiency versus comprehension. Here we define a new measure of silent sentence reading efficiency (SRE) and explore key aspects of item development for an unproctored, online SRE assessment (ROAR-SRE). In doing so, we set forth an argument for developing sentences that are simple assertions, with an unambiguous answer, requiring minimal background knowledge and vocabulary. We then run a large-scale validation study to document convergent validity between ROAR-SRE and other measures of reading. Finally we validate the reliability and accuracy of using artificial intelligence (AI) to generate matched test forms. We find that a short, one-minute SRE assessment is highly correlated with other reading measures and has exceptional reliability. Moreover, AI can automatically generate test forms that are matched to manually-authored test forms. Together these results highlight the potential for regular screening and progress monitoring at scale with ROAR-SRE.