Answers are grounded in this paper and cite their sources.
Demonstrates the effectiveness of interview-grounded digital twins, suggesting new evaluation methods.
Project DOPPLER asks a narrow question: can a language model that has read a person's earlier public interviews answer a later, unseen interview question the way that person actually did, and how would you check? The methods paper reports a negative result. Forced choice, the usual way this kind of twin is scored, turned out to be solvable without knowing the person at all. A person-blind scorer, told nothing about the subject, picked the correct answer 17 of 17, 10 of 10, 15 of 15, and 8 of 8 times across four different distractor constructions. A kill rule written before the data fired, and the instrument was replaced with open-ended generation. The results paper reports the confirmatory finding under the replacement instrument, scored on two independent channels. On 88 subjects drawn from MediaSum, a public archive of 463,596 NPR and CNN interview transcripts, a twin grounded in a person's own earlier interviews beat an identically built twin grounded in a different person's interviews by 0.075 embedding cosine and 12.1 stance-match points, both significant. The registered magnitude bar was met on that contrast and missed on the own-versus-zero-information contrast the frozen text names; both are reported in full. The study was preregistered, splits are strictly chronological, and the primary metric throughout is lift over a zero-information baseline rather than raw accuracy. This deposit contains a consolidated preprint, a plain-language article, and the two underlying technical papers.
No takes yet. Share an insight, caveat, or question.
Umar Aslam (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: