Use AI feedback as a starting point: inspect rubric matches, evidence and limits before confirming marks.
Evidence-led AI grading compares readable student work against explicit rubric criteria and explains the evidence behind each proposed mark. It does not transfer the final decision to an AI system. Teachers should inspect the cited work, confirm benchmark matches and correct unsupported conclusions before sharing feedback.
Check that every rubric row appears once, its maximum points match the source and its selected benchmark descriptor is preserved. Read the actual cited passage or page. An evidence reference is a location to verify, not proof that the interpretation is correct.
Progress checkpoints, participation and physical production often require evidence that an uploaded PDF cannot supply. Keep those marks provisional and add teacher observations. Missing evidence must not become invented evidence, automatic full credit or an automatic accusation.
The course-first rubric workflow can support readable science reports, mathematics explanations and programming documentation. Specialist correctness still needs teacher review. Running code, validating CAD geometry and assessing video are separate planned capabilities, not existing promises.
Product capabilities and limits are recorded in our roadmap. Questions? Email the team.