Elicitation & personalization

Getting one person's actual judgment into a machine, and measuring honestly whether it got there.

Single-subject (N-of-1) methodology, worked out in practice: how to mine a person's real past decisions out of years of transcripts; how to keep a held-out set of those decisions hidden from the model being tested, so it can be checked against choices it never saw; and how to grade a stand-in persona's answers without letting the same system grade its own test. It favors what someone actually chose over what they say they would choose, checks whether presenting a decision as a fictional scenario changes the answer, and watches for cases where the person being modelled knew something the grader did not. What it is working toward is a stand-in current enough to be trusted with a real decision, and a number that says whether it is.