Audience segments drawn from what people actually write — trait distributions, segment differences with their uncertainty, and a clear line between what the data shows and what we think it means.
Echo Persona counts twelve behavioural language markers (bilingual lexicons, diacritics folded) in each text and reports, per segment, the share of texts carrying each marker with a Wilson 95% interval. A trait becomes a segment's signature only when the Newcombe interval of its difference from the rest of the corpus excludes zero. When no segment field is given, segments are discovered with k-means (k = 3–5 chosen by silhouette) on rarity-weighted, normalised marker vectors; their stability is the share of co-assigned pairs that stay together across five bootstrap re-clusterings. Segments under 20 texts are displayed but never interpreted.
Limits: markers are what people wrote, not who they are. Interpretations are hypotheses about a segment's motive, each printed with the observation that would falsify it. The tool does not create individual profiles and refuses protected attributes as segment variables. An earlier version of Persona was withdrawn for review; this rebuild answers that review by working only at segment level.