Privacy and fairness

Measuring contribution responsibly means drawing a hard line: enough visibility for fairness, no more. Dwixel’s privacy model is a feature, not a footnote, and it is enforced in the product.

Privacy and access, by design

Instructors can review the group’s work as part of oversight, and see who contributed what. What Dwixel does not do is turn that into covert profiling: the contribution measurement is deterministic and explainable, students always see their own data, and peer review is collected confidentially. A review of 77 studies groups the ethics of educational trace data into four areas: privacy and consent, validity and integrity, ethical decision making, and governance. ¹ Access control and transparency speak to the first and the last. Whether a count of words validly represents a contribution is a separate question, and not one that being open about the method settles.

What a student is shown about themselves

A measure that decides part of a mark and cannot be questioned is not a measure, it is a verdict. So the student side is built to be argued with: their own share against an even one, how far peer ratings can move their mark stated in marks rather than as a weighting nobody can interpret, what the measure cannot see, and a plain list of everything worked out about them.

The cap on how far peer ratings can move a mark is set per module and published to the student rather than kept in a settings panel, because the reason a cap reduces the pressure students report is that they know what it is. ² There is a floor as well as a ceiling, and Dwixel warns about both: set the weight low enough and students stop treating the exercise as consequential at all, which does not make peer assessment gentle so much as pointless.

The list of what Dwixel works out is split in two, and the split matters. A student cannot opt out of being assessed, so presenting the measurement as a choice would be a consent that cannot be refused. That part is a notice. Only the genuinely refusable uses — appearing in institution-wide dashboards, and anonymised research — are consent, and refusing them changes what happens. Consent also has to be ongoing rather than given once at enrolment, because the value of trace data becomes apparent only after it has been analysed, not beforehand. ¹ When the list changes, the student is asked again and shown only what is new.

Confidential peer review, done carefully

Peer assessment is a useful second signal, and Dwixel records who wrote each rating while never showing it to the person being rated. That is partial anonymity rather than blanket confidentiality, and it is what Yang recommends by name: evaluations turned in with the rater’s name attached, blinded before they reach the person evaluated, so the honesty of anonymous feedback survives while the rating stays attributable to staff. ² WebPA has shipped the same mechanism for fifteen years. ³ The research here is candid rather than settled. Peer scores inflate once a grade is attached, most of all for underperforming students, ² and in a study of students rating one peer’s individual work, a rubric improved how closely peer marks tracked an expert’s while the friendship bias came out larger with the rubric than without it. ⁴ Agreement with teacher marks is adequate in most studies and unacceptably low in a substantial minority, and peer assessment of individual contributions to a group project is one of the applications named in that minority. ⁵

Two findings run in the other direction and are worth stating. Peer—teacher agreement is highest of all for academic processes, meaning participation in group activities, which is exactly what is being judged here. And peers do not systematically mark above or below a teacher: setting one outlying study aside, the average gap between them is effectively zero. ⁶ Group size is not the lever it is often assumed to be, though. Agreement got weaker as the number of raters rose, and single raters were no less reliable than groups. ⁶ So the peer score weights a mark in Dwixel. It never sets one.

Access control and recovery

• ·Access is enforced by row-level security in the database, not by interface convention.
• ·A student’s work is theirs, and their contribution data is shown back to them in the same view their instructor gets.
• ·Documents and presentations are backed up automatically, with attributed version history, so accidental loss is recoverable.