This whitepaper explores whether the history a physician leaves in a clinical search engine can be turned into a portrait that describes that specific person rather than a physician type. Four mutually blind language-model agents read each history from the perspective of a clinician, an educator, an information-behavior researcher and a skeptic, producing portraits with a median of about 10,700 tokens. The portraits were tested on 339 physicians and 34,241 searches from Arkangel AI with three evaluations: identification, specialty and future query generation.
When each history was split in two, the portrait from one half matched the same physician's other half in 78.4% of cases (chance 0.3%), but simply concatenating the raw search text reached 85.0% and beat the portrait in every volume stratum. The portrait identified the self-declared specialty in 49.9% of physicians versus 41.3% from raw text, an advantage that came from generalists, and in query generation it ranked the correct physician higher for 157 physicians versus 96 (p < 0.0001), although the top-1 difference was not significant (p = 0.17). Agreement between portraits from different halves was limited (median ICC 0.496 across 22 dimensions), and a later code review found a train-test overlap affecting 3.7% of the holdout. All findings are exploratory; it is a whitepaper, not peer reviewed.
