A Governed Workflow for LLM-Based Formative Assessment with Score Stabilization

Abstract

Open-ended formative assessment requires records that teachers can interpret and trust over time, yet large language model outputs remain variable across repeated calls. This study develops a governed workflow for LLM-based formative assessment of student core competencies. The method combines four competence dimensions, a JSON-only response contract, backend validation and normalization, threshold-based repeated inference with closest-pair aggregation, deterministic post-processing, and teacher review before publication. It is implemented in a classroom platform that stores traceable assessment records and supports follow-up reporting. Technical checks showed end-to-end execution and preservation of structured result objects with ensemble metadata. In a one-term pilot, 147 experimental and 145 control records met the strict confidence-analysis criteria. The confidence distribution differed significantly between groups, and within the high-confidence A-level subset the experimental group showed a larger later-stage reading gain than the control group (10.52 vs. 3.51, p = 0.0029), whereas writing differences were not stable across filtering choices. These results do not establish causal effects, but they indicate that governed processing can turn variable LLM outputs into reviewable records for classroom use.