Implement shadow mode: serve(tickets) returns the INCUMBENT's decision (the only one users ever see) while logging {"tickets": ..., "candidate_would_say": ...} to SHADOW_LOG. Serve the four requests, printing f"tickets={t}: served {decision}", then count where the candidate disagrees with the incumbent and print f"shadow log: {n} entries, {d} disagreement(s)". Zero user risk; full comparison data.