DEMOS · a synthetic audience scored against reality

THE CUSTOMER PARLIAMENT

~150 synthetic customers bred from first-party telemetry, each tagged with the evidence it was bred from. Any artifact can summon a jury; every verdict is a falsifiable prediction reconciled against real Umami/Stripe outcomes and Brier-scored. Chronically wrong personas are culled and re-bred.
loading…

Population by niche

Reliability — is the jury's confidence earned?

Each resolved prediction is binned by the probability the jury assigned, then plotted as predicted (x) vs observed conversion rate (y). A perfectly-calibrated panel sits on the dashed diagonal. Points above the line = the jury was too pessimistic (under-confident); below = over-confident. Dot area ∝ how many predictions fell in that bin.

Calibration ledger — Brier + log-loss

Brier: 0=perfect · 0.25=coin · 1=worst. Log-loss punishes confident misses (coin≈0.693). Ranked worst-log-loss first; ◈ flagged = cull candidates.
brierloglossnpersonanichestatus

Lineage — bred children

childnicheparents

Latest jury