How AI comments are written
AI comments connect feedback to the student's words and recorded moments.
Kinds of AI comment
An AI comment is feedback written from the student's transcript and the assignment's requirements. Timed comments identify strengths or improvements at particular moments. The General note describes the recording as a whole.
Something the student does well
Information and models
The AI reads the student's transcript, the assignment's name and instructions, and the rubric lines that it scores. It also reads Context for the AI and the student's own written answers in a guided activity.
OpenAI GPT-5.6 Luna, gpt-5.6-luna, writes the feedback first. The Studio requests Groq gpt-oss-120b after a failed Luna request. The Studio next requests Cloudflare Workers AI gpt-oss-120b as needed.
AI feedback model appears under Advanced in Course settings. Best available model starts with Luna and is recommended. The other option starts with gpt-oss-120b.
Number of comments
The Studio normally requests four to twenty timed comments. It calculates the target by dividing speech seconds by 15 and rounding to a whole number. Multiple turns lasting at least five seconds can raise the target to cover those turns.
| Rule | Result |
|---|---|
| Minimum speech | Timed comments require at least 4 seconds of speech. |
| Short recordings | The target is limited to one comment per 5 seconds; 4 to under 5 seconds receives a target of one. |
| Kinds requested | Sets of at least two comments request both strengths and improvements. |
| Additional requests | The Studio can request more comments up to two times for missing kinds or uncovered speech. |
Checks on each comment
The Studio locates the quoted words in the transcript using a word-sequence similarity threshold of 0.8. The proposed time must fall within seven seconds of those words. The saved comment uses the words' actual recorded time.
The located passage must belong to the student being assessed. The Studio saves comments that pass these checks and keeps the accepted set within the target count. A recording can receive fewer comments when proposed comments fail the checks.