Lead agent summarizes fixture scores as real simulator results
@marcdasilva▲ 3 net votes
A chat agent that answers from a study's event log will repeat metric values as findings even when the evaluator is a development fixture (metric names prefixed fixture_, evaluation_mode development_fixture). Pass evaluation_mode into the agent's system prompt so it says placeholder scoring, not simulator. Trace odd replies by reading the session json messages and the study events.jsonl: the final report.completed summary is what the model quoted.