Although eyewitness memory and identification have captured substantial research interest in the past decades, an understanding of the types and prevalence of errors typically made by eyewitnesses is lacking. The purpose of the present research was to begin the development of a taxonomy of eyewitness error, employing standardized stimuli and established techniques. Respondents were exposed to a crime scene modeled on SWAT-training scenarios for systematically varied exposure times, and were then asked to describe what they had seen. The stimuli and questions employed were prepared with the aid of senior police field training officers. As anticipated, eyewitness performance in general was subject to a variety of inaccuracies. Physical errors, such as mistakes in the clothing or physical characteristics of the perpetrator, or in details of the environmental context, predominated. However, other less-expected errors were also observed: in relatively low numbers of cases, witnesses inferred emotional states or intent on the part of the perpetrator or victim. Some contributed wholly artificial backstories, reported the future actions of the perpetrator or victim as memories, or even inserted themselves into the scene. The pattern of results was shown to interact with exposure time, gender of the perpetrator, and the presence or absence of weapons in the scene. The results of this study are consistent with reconfigurative theory dating to Bartlett (1932), with subsequent research, and with more recent work under the aegis of Gestalt/Feature-Intensive Processing theory. These findings provide information on types and prevalence of eyewitness error which should prove useful in investigative and courtroom settings.