Assignments & grading
A rubric, so two markers agree
Assessment & Outcomes
Online exams from a question bank: papers with sections, attempts, an anonymised marking queue, and item analysis that names the problem question.
assess · 42 API routes
Reused across papers and terms, which is the whole reason to have one.
Set to a course, a batch or a person.
One learner run, with its own state machine and its own audit trail.
It drops anything already marked, so two markers cannot collide.
Which questions discriminated and which were noise.
What a cooperating browser reported, shown to a human.
How it actually works
No camera, no automatic verdict, nothing that acts on its own. The browser reports what it can, the candidate is told while it happens, and a human marker sees the signals with the caveat on the same screen. An automated cheating verdict from a browser event is a false accusation waiting for a lawyer.
A sample. Each is a real key in the catalogue, described for a school administrator rather than an engineer.
assess.question.viewassess.assessment.publishassess.attempt.evaluateassess.result.viewThe shipped role templates that hold it by default. Each is a copy you own and can edit — nothing in the code reads a role's name.
Naming these here is cheaper for both of us than discovering them in week six of a pilot.
Naming these is cheaper for both of us than discovering them in week six.
A rubric, so two markers agree
Verifiable by a stranger
A hundred seats, twenty-five courses and no conversation. If you would rather see it driven by someone who knows it, that takes twenty minutes.
Or open the sandbox (opens in a new tab) and click around first. No form.