Research
The software is also the instrument
Mathazel is the platform for a postgraduate study of how pre-service teachers transform routine textbook problems into modelling tasks while working with an AI assistant. The same codebase runs the study and would run the product; the boundary between them is enforced in the database, not by convention.
The framework is Maass (2010) for what makes a task a modelling task, and Blum & Leiss (2007) for the cycle a solver travels. Both process and product are analysed: the finished task is rated blind against seven criteria, and the interaction record shows how it was arrived at.
Two decisions that shape the thesis, not just the code
Hazel drafts freely on request. Refusing would waste the teacher's time, so the construct being studied is "teachers working with an AI design assistant", not "AI as co-designer". In exchange, deference is measured rather than assumed: the overlap between what the assistant wrote and what the teacher committed is computed on five-word shingles, so lifted text scores 1.00 and original text scores 0.00.
No identifying detail enters the system. Participants are known only by a code on a slip. The consent form carrying a name stays on paper with the researcher, and a study session cannot hold an account id, that is a constraint on the table, not a rule someone has to remember.
The platform has a second side, and the study does not use it
A task designed here can be published as a link and worked by a solver, who moves through the seven transitions of the cycle and marks where they stall. That is the product half of the codebase, and it exists because the question a designer most needs answered is not whether a task is good but where it actually bites.
A study participant's task can never be published. The route refuses it before it writes anything, and that refusal is not a configuration flag. A participant consented to a supervised sitting, not to their work being handed to strangers on a URL, and changing that would be an ethics amendment rather than a setting. It is the same class of decision as the flag that keeps participant work out of the shared retrieval pool.
Solvers are not participants either. They arrive on a link, they are asked for nothing that identifies them, and their sessions are excluded from the study export by mode. Two self-report instruments exist for product use, a designer scoring their own task and a short set of questions for solvers, and both stay switched off in study mode until their wording has been through ethics with everything else.
Where it stands
Data collection has not begun. The wording participants will see is still going through ethics, and the retrieval library, 25 curated items spanning school, undergraduate and postgraduate material, is loaded and logged so that a library which grows during collection becomes a covariate the analysis can control for rather than an invisible confound.