Models should be benchmarked against the application's own reaction deadline rather than relying on general leaderboard scores; latency should be treated as a first-principles constraint in model selection rather than an after-the-fact optimization metric.