Grounded personas & predicted behavior
Personas now ground in real source material with claim-level provenance, sessions cite the exact source lines, and outputs move from opinions to concrete, falsifiable behavioral predictions.
The biggest credibility gap was personas reasoning from prompts rather than real material. That gap closes today, end to end: bring real evidence in, trace every claim back to it, and get falsifiable predictions out.
Real material in
- Corpus ingestion: interview transcripts, support tickets, reviews and survey verbatims become deduped, citable chunks — messy input handled mechanically, no model in the loop.
- Grounded persona authoring: create a persona primarily from uploaded material, or refresh one against it — with provenance recording which chunks shaped which trait.
- Evidence files: attach screenshots, PDFs and images to a project; image evidence is viewed as real pixels before any reaction is authored, and councils carry every asset in the room.
Falsifiable predictions out
Sessions, councils and syntheses now emit predicted behaviors — "abandons at the price reveal, likelihood 0.7" — each on a canonical likelihood scale and tied to its evidence. The segment roll-up shows where a cohort drops off, and recurring predictions promote into hypotheses scored against reality.
Improvements
Every persona context carries its task-relevant source chunks with a citation contract (evidence refs with quotes).
Evidence assets and grounded material ride each participant’s context automatically.
Research projects and evidence binaries persist through the portable snapshot round-trip.