What I did
I marked one real cohort of three typed submissions against a five-criterion rubric, from start to finish in Feedbacker. I imported the submissions from the marking platform's bulk download, anonymised and approved each one, asked the AI for proposed levels, marked every criterion myself, drafted feedback from my own marks and comments, checked it, approved it and exported it. This is public, so I can't say which institution, module or students: the cohort is described only by its shape.
What I found
- The AI proposed a level for 12 of my 15 criterion judgements, and my level matched all 12. I marked with the proposals showing, so that is not a test of whether the AI agrees with an independent marker.
- The feedback needed drafting again. All 18 pieces of feedback (five criteria and an overall comment for each of three students) started from an AI draft. The first drafts sometimes came back cut short, or with a stray quote, so I revised the instructions and drafted them again instead of editing by hand. The history holds 27 earlier versions.
- I accepted four flags, and all four were false alarms. Each was a flag for naming another mark, and each was a figure I had quoted from the student's own work, not a mark. Flags I cleared by changing or redrafting the text weren't recorded, so I don't know how many were raised in all.
- Nothing was released without my approval. All three submissions were approved before export, and only approved work was exported.
- I didn't time it. I judged that a timed run would add little. My verdict is that it works well.
What it changed
Marking a real cohort found problems the made-up coursework hadn't: how weights were labelled, drafting instructions that produced cut-off replies, and unclear outcomes on export. I fixed each before I finished.
The limits
This is one educator and three submissions. I marked openly, so there is no blind comparison, and I didn't measure the time saved. It showed no need yet for several markers to coordinate.