Open-Source 'Assessment-First' Course Uses LLM Rubric Grading, Not Chatbots

Developer Michael Tuszynski has released doerkit, an open-source statistics course built around the idea that LLM-powered rubric grading and spaced cumulative review are more effective than AI chatbots for improving student outcomes. The system grades written answers against predefined rubric criteria in roughly one second, returning specific feedback and partial credit rather than simply assigning scores. Two MIT-licensed repositories — doerkit and rubric-bench — have been published, with the latter providing regression and adversarial testing for any LLM-based grader. Tuszynski acknowledges significant gaps between the current demo and a deployable product, including missing LMS integration, multi-tenancy, FERPA compliance, and validation against human raters. The project's core findings note that grader warmth and severity are independently adjustable, that frontier models resist prompt injection better than cheaper alternatives, and that cumulative review showed the largest effect size in the underlying source study.