Builds likelihood questions for each candidate note, keyed n1, n2, etc.
Each question quotes its own note. Question keys are labels for code, not instructions to the
model, and a native decision driver (Jev) may answer each question separately against the state,
so a question that only named its key would ask the same thing about every note. Measured: with
the note left out of the question, Jev's answers were at chance (AUC 0.51).
Builds likelihood questions for each candidate note, keyed n1, n2, etc.
Each question quotes its own note. Question keys are labels for code, not instructions to the model, and a native decision driver (Jev) may answer each question separately against the state, so a question that only named its key would ask the same thing about every note. Measured: with the note left out of the question, Jev's answers were at chance (AUC 0.51).