
82 of 84 marking verdicts matched the answer key in testing
An agent for one faculty member that works one week's assessment: it marks every answer against the rubric, drafts the feedback, and holds every line until the faculty member approves it.
The test, in figures
- 20 of 21
- Graded answers whose outcome matched the answer key
- 82 of 84
- Criterion verdicts matching the key
- 1
- Dangerous error: a partial answer the agent passed
- 7.6
- Seconds per graded answer, median
The problem
The Chancellor of a Dubai government university said on the record that every faculty member now has an artificial intelligence agent, for the content, the assessment and the feedback. He said feedback is where higher education suffers. He gave no figure for agreement, the share of the agent's marks that match a person's.
The build
For every answer the agent reads the rubric and returns a pass or a fail on each criterion, and the code turns those verdicts into the mark: pass, partial or fail. For every answer that does not pass it drafts the feedback, with a worked example. After the last answer it writes a one-paragraph digest for the teacher and proposes one change to the material. Three rules refer an answer to the faculty member unmarked before any model sees it.
The test
On 26 September 2026, 24 invented answers ran one at a time. Three were referred by the rules before any model call, as the key said, and no graded answer was referred by mistake. The outcome matched the answer key on 20 of 21 graded answers, with one dangerous error. The whole class of 24 took 151.2 seconds.
What a person still decides
Nothing reaches a student until the faculty member approves the line. The release count in the frozen run is zero.
Read the full paper
Issue 1, Rev A, September 2026. It gives the problem, the design and the test, with every source and the file each figure comes from. Read the full paper as a PDF.