Historical figures, reconstructed

Sources & method

How it is made

• An AI model (Claude) acts as a historian rating someone in the third person. It reads a dossier of cited sources: the person's English Wikipedia article and, where we added them, public-domain excerpts of their own writing or of people who knew them.
• It answers the same questionnaire you can take on Deep Person, about 1,900 agree / neither / disagree statements, judging the person against people of their own time and station.
• Items go in batches of 40, and every item is rated 3 times in a different order. The most common answer wins; a tie counts as “neither”.
• When the sources give no basis for an item, the rater says so. Those items are left out of the scores, not guessed.
• Scores are computed the same way as your own profile. Only traits suitable for a public page are shown, and the quotes behind each one are listed under “Why”.
• It is a reading of the record, not the person speaking. You cannot chat with a figure, and nothing here is presented as their own words except the quoted sources.

How reliable it is

• Repeatability: two independent ratings of the same figure from the same sources agreed at r = 0.98 across traits (2 figures rated twice), and 0% of items flipped from agree to disagree.
• Yes-saying bias: 0.00 on average (0 means none). 4 of 12 runs were slightly outside our ±0.10 target.
• Agreement with experts: Openness correlated r = 0.94 with published expert ratings of 8 US presidents (Simonton, 2006). The other four Big Five domains have not been checked yet.

These are small samples, measured November 2026. Repeatability shows the rater is consistent, not that it is right.

Figures is experimental. More figures are coming, and the method may change as we measure it.