Compare runs without fooling yourself

Answer whether the observed change supports the customer's decision. Stay read-only. Do not start a hill-climb, create variants or launch confirmation calls unless explicitly requested within a budget.

Freeze the comparison

Confirm organization/workspace, API environment, baseline and candidate run IDs, the intended change, primary criterion and acceptable regressions. Read each run and its individual conversations, selected metric outputs and relevant recordings. Use the installed CLI's runs get, simulated-conversations list --run-id, get, metrics and metric-detail commands. Check ok and paginate through the public API if the CLI can't prove a complete list.