Evaluating alerts, backtesting with opus4.8 and gpt5.5, going back and forth for 4 rounds, results: 1️⃣ Accuracy: 5x stronger than random; 2️⃣ Coverage: Detection layer basically catches everything, almost no false positives when filtering signals; 3️⃣ Timeliness: On average, there's still 21% room for movement after detection. As a short-term alert capture, it's pretty much optimized. Wizz Alert: t.me/wizzalert