Understanding AI Confidence Scores and Review Flags

Every question E-Porikkha grades comes with a confidence score — a number from 0 to 100 that reflects how certain the AI is that the awarded mark is correct. Understanding this score helps you focus your review time on the answers that actually need a second look.

The grading model compares the student's extracted answer to the question's model answer across several dimensions: factual correctness, key concepts present, and how well the reasoning lines up with the marks scheme. The confidence score reflects how certain it is about the mark it gave.

A high score generally means a clear match or a clear miss. Lower scores indicate partial matches, ambiguous phrasing, illegible handwriting affecting part of the answer, or answers that are correct in spirit but worded very differently from the model answer.

A paper's overall confidence is the average across all its questions. When that average is below your school's auto-approval threshold (88 by default), the paper is marked Ready for Review instead of being auto-approved, and appears in your Papers queue for you to check.

Papers are also flagged when the AI detects a potential contradiction (e.g., the student says both X and not-X), or when part of the script was illegible and the AI couldn't verify the answer at all.

The auto-approval threshold is a school-wide setting. Raise it (e.g. to 95) for high-stakes exams where you want to manually check every borderline case. Lower it (e.g. to 75) for routine class tests where speed matters more than reviewing every paper by hand.

In the grading studio, click the copilot icon on any flagged answer. The drawer opens a chat where you can ask the AI to justify its score, compare to a different part of the model answer, or propose a revised mark. You confirm or reject the proposal — the AI never changes marks without your explicit approval.