New Agent Picks re-tests decide Improved, Declined or No clear change with the 90% interval on the difference between the re-test's pick rate and its baseline, the check Head-to-head uses. Their ranges are 90%, and a re-test stops early once that change is clear.
New tests decide a loss the same way: a cell is a loss only when a competing tool or Built its own was clearly picked more often than your tool at 90%. A cell your tool leads only when it was clearly picked more often than everything else.
Earlier results are not recalculated. Re-tests already running finish with the method they started with, and their results, like every earlier re-test result, keep 95% ranges. Earlier tests keep their losses and Actions. Results shows Method changed when a list includes both.
Over API v2 and MCP, each Agent Picks Action's result and retest carry method: v2, or v1 for a re-test started before this change, whose ranges are 95%. The Outcomes ledger's result_evidence.agent_picks and Timeline outcome_measured events (agentPicksMethod) carry it too. Cells carry decision_method beside is_loss and own_led, and the summary's test carries it for the shown test. A missing value is v1.