If you have never sat on a panel, the mental model is usually wrong. People imagine a trial. It is closer to a careful conversation with a written record, run by people who are acutely aware that they can damage someone's degree.
That framing explains almost everything about what works.
The standard is lower than a court and the caution is higher
Most institutions decide on the balance of probabilities rather than beyond reasonable doubt. In practice panels are conservative anyway, because the consequences are severe, the student is usually young, and appeal processes have teeth. A case that rests on one contested number tends to end in no finding, and reasonably so.
What collapses
A detector percentage on its own. It is a claim the panel cannot inspect, produced by a system the panel cannot question, about a document rather than a person. Many institutions now say in writing that a score alone is not sufficient grounds.
"It reads like AI." An experienced instructor's instinct is often right and is not evidence. Once it is written into a case file it invites the obvious question: what specifically, and how do you distinguish that from a careful student writing formally?
Style comparison against earlier work, used alone. It gets closer, but improvement over a term is normal and expected, and a panel will say so.
The pattern is consistent. Anything the student cannot meaningfully answer tends not to survive, because a process that cannot be answered is not a process.
What holds
Fabricated sources. A citation that does not exist, or exists and says something else, is checkable by everyone in the room. It is the single most decisive piece of evidence in this category of case.
Document and version history. Creation timestamps, paste events, a 2,000 word jump in one revision. Concrete, inspectable, and the student can offer an explanation that either fits or does not.
A conversation about the work. A student who cannot explain the method in their own submission, cannot say why they chose it, and cannot answer a follow-up about a specific paragraph has given the panel something it can weigh directly. Crucially, they have also had a fair chance to demonstrate the opposite.
Anything with a record. Panels trust artefacts they can review after the fact more than they trust anyone's recollection, including the instructor's.
The design lesson
Notice that the items that hold share one property: the student can engage with them. Fabricated sources can be explained if there is an explanation. A version history can be contextualised. A conversation can go well.
This is not a coincidence and it is worth carrying into how you design assessment in the first place. Evidence that a student can answer is evidence a panel can use. Evidence that a student cannot answer puts everyone in a room arguing about credibility, which is the worst version of this process for the honest student and, uncomfortably, the best version for the dishonest one.
Related reading: the students AI detectors get wrong.