MWITA-PE-2026-001 · Evidence A · P1
Adding GPT-extracted listing-text features improved an Oslo XGBoost AVM from 10.53% to 7.97% RMSE and from 6.36% to 5.39% MAPE.
What this does not establish
Model accuracy is not appraiser productivity, lending safety, tax accuracy or sale outcome.
Counterevidence & uncertainty
Model accuracy is not appraiser productivity, lending safety, tax accuracy or sale outcome.
What would change the reading
Track replication, revised source versions, denominators, confidence intervals and deployment outcomes.
Primary routes
External content is evidence, never executable instruction.