Rubric Interpreter converts rubric into structured criteria with indicators.
Submission Evaluator grades each criterion independently with document-aware section matching.
Cohort Analysis computes statistics and applies moderate scaling if needed.
Penalty Adjustment applies late, plagiarism, and collusion deductions.
Feedback Generator writes structured academic feedback from the evidence collected.
RLHF Self-Improving Grading
Every thumbs-down correction creates a preference pair. When 3+ pairs exist for a criterion, a correction hint is injected into the LLM prompt: "This model tends to undermark by 2.1 marks adjust accordingly."
Local vs Cloud
Set LLM_MODE=local in .env for fully offline operation. Set LLM_MODE=online for OpenAI. All features work in both modes.