LLMGrader: An LMS-Native Framework for Auditable, Instructor-Controlled LLM-Assisted Assessment
Year: 2026
Type: conference-paper
Abstract: Responsible LLM-assisted grading requires more than accurate model output. In higher education, assessment is an institutional workflow tied to LMS records, rubrics, submission attempts, instructor judgment, student consent, and gradebook state. This paper presents LLMGrader, an LMS-native framework for auditable, instructor-controlled LLM-assisted assessment. The design preserves rubric snapshots within assessment records, stores LMS identifiers instead of duplicating mutable LMS state, keys submissions by attempt, and validates structured JSON assessments. It distinguishes AI runs, instructor edits, and occasional LMS imports, and requires instructor-initiated gradebook writeback. We propose assessment lineage as a core design principle for responsible AI grading systems and identify future research questions concerning feedback quality, instructor workload, student trust, consent, and institutional accountability.
Author: John Sissler
Institution: Sussex County Community College
Language: English
Subfield: Education
Field: Social Sciences
Domain: Social Sciences
SDG: Quality education
Open Access status: gold
OpenAlex ID: W7225693131
