Intuition Labs · φ research · 09 · mixed · 23 may 2026
production traffic is the eval set
average the model’s confidence across production streams. when it drops, the model has drifted.
folded into the-signal · shipped as the alarm the loop raises on itself and the diagnosis that separates real capability loss from the wrong model on the right problem; read note 30.
~48× faster detection
alert ⇔ rolling_mean(φ_auc, 7d) drops > 2σ · drift signal