Research lines
Four lines. They're not equally well supported, and I've tried to say where each one is thin rather than level them out.
Construct validity in machine judgement
Whether a test measures what it says it measures, or just the surface features that travel with it. My own probe is the worked example: the control killed the headline finding and left something smaller behind. One dataset, with intervals. The analyses were exploratory, and there's a blind expert validation packet I built and still haven't run.
Evidence governance and denominator authorship
Who writes the list a system is graded against, and whether the system could see it. If that list comes from the system's own account of what it owed, then leaving something out removes it from both sides of the fraction and the score doesn't move. I have engineering evidence on three separate axes with mutation tests behind it, and one real case from the field. The independence is procedural, though: I can show the reference didn't move during grading, not that it wasn't written from the same material it grades.
Human and machine decision authority
Which parts of a decision should be written by a model, derived mechanically, or confirmed by a person — and what a legal standard of care quietly assumes about whoever it is measuring. Article 14 of the EU AI Act writes down the capacities that assumption was always making, then leans on capacities the technology it regulates wears away. This line is analysis. The empirical work underneath it is other people's.
Roman law and the standard of care
The bonus et diligens pater familias was never a fixed measure. It's a semantic field, and it bends with the internal logic of whatever legal relation it gets applied to. The tradition is useful here because it kept two names for two things a trained model runs together: an ideal of how one ought to behave, and a record of how people did. This is my thesis work, and the paper derived from it isn't finished.