How Discentik grades your work

Practice activities in a course are graded by AI against the requirements the course
author wrote for that activity. This page explains what is graded, how the result is
decided, and what Discentik does to make the same work receive the same grade.

What gets graded

Each activity grades one thing, chosen by the course author:

• Your prompt: the instructions you gave the AI. A good answer from the AI does
  not count if your prompt did not ask for it.

• The AI response: the answer you produced with AI, read together with your prompt.

• Your document: in a Canvas activity, the current document. Your chat messages
  are not part of the grade.

The grader only sees that submission. Actions it cannot see, such as copying, pasting
or how many times you revised, are never graded.

A checklist, not an impression

Every activity has a short checklist of requirements, such as "Both dates included".
You see the same labels in your feedback. The grader marks each requirement:

• Met: your work fully does it.

• Partly met: you attempted it, but it is incomplete or unclear.

• Missing: your work does not address it yet.

For every requirement it marks as met, the grader must quote the exact words from your
work that show it. Discentik then checks that the quote really appears in your
submission. A requirement whose quote cannot be found is not counted as met.

How the result is decided

The result is calculated from the checklist, not chosen by the AI:

• Passed when every requirement is met.

• Your score counts each met requirement fully and each partly met requirement as
  half. For example, 2 met and 1 partly met out of 4 gives 63.

• Almost there means you did not pass yet but scored 50 or more. Needs work
  means a score below 50.

Some requirements can be checked exactly, for example that two dates are stated. These
are checked by fixed rules, and a rule result always takes priority over the AI.

Making grades consistent

AI can word the same judgement differently from one run to the next, so Discentik
grades in a way that keeps results stable:

• Grading always uses the same AI model with its most predictable settings and a
  strict answer format. Discentik checks how consistently a model grades before it is
  used, and changes the model only deliberately.

• If any requirement looks borderline (partly met, or met without a quote that can be
  found in your work), the work is graded twice more with the checklist in a different
  order, and each requirement takes the majority result. Ties go to the stricter result.

• Course authors see which requirements were judged differently across runs when they
  test a course, so unclear requirements can be reworded.

Grading is consistent, not perfect. Rewording your work can change a borderline
requirement, and the AI can still misread something. If you think a result is wrong,
you can ask a person to review it.

Reading your feedback

Your feedback opens in Your review, which names the attempt it belongs to. When
an activity is not passed yet and the review is closed, its button shows how many
requirements you met, for example 3 of 4 met. The note at the bottom of the review
is a reminder that AI checked your work against the course checklist.

When an activity is not passed yet, the feedback starts with Your next step: the
single change that will help most. Below it, What to fix lists each requirement
that is missing or partly met, with the words from your work the grader looked at and
a suggestion to try. Select Show what you did well to see the requirements you
already meet.

When you pass, the feedback shows What worked and, sometimes, optional ideas under
Make it even stronger. You can reopen it with View feedback.

Attempt history shows the score of each attempt, so you can see your progress.

What is stored

Discentik stores the result of each attempt: pass or fail, the score and which
requirements were met. The quotes and suggestions in your feedback are stored with
the same rules as the rest of your attempt text. If your organisation does not keep
attempt text, your organisation's reports show results without quotes from your work.

For course authors

Requirements that are specific and observable grade most consistently:

• Write one requirement per line, in the learner's terms, for example "Subject line
  requested" rather than "The learner writes well".

• Use How the grader checks this to say exactly what must appear.

• Keep to about 3 to 6 requirements per activity.

• Do not grade actions the grader cannot see, such as copying or revising.

The course editor shows a hint next to any requirement that is vague, checks more than
one thing or describes an invisible action. Use Test with a strong and a weak
answer before publishing.