Been building round handicap models for CS2 for a while now, and the #1 mistake I see people make when they start tracking their reads isn't the model — it's the scoreboard they use to judge it.

Most people track win rate on flagged rounds. That's noise. Here's why, and what to track instead:

1. Calibration error, not win rate.

If you flag a round as a 70% handicap toward T-side and it hits 68% of the time across a real sample, that's a great model — even if you'd have "lost" plenty of individual reads. Win rate on a single game tells you nothing. Track how close your predicted probability is to the realized frequency over 50+ rounds. That's calibration, and it's the only thing that tells you if your scanner is actually reading signal.

2. Edge decay across a half.

Your round handicap read is almost never static — eco rounds, force-buys, and momentum swings change the number bucket by bucket. If you're only snapshotting your handicap at pistol round and holding it constant, you're blind to the biggest edge swings in the game (round 3-6 force buy reads are where most of the value is hiding).

3. Sample-adjusted confidence, not raw hit count.

"I hit 8 of my last 10 flagged rounds" means almost nothing at n=10. Track your rolling confidence interval, not your streak. A specialist who's honest about variance will out-earn one who's chasing hot streaks, every time.

If you're building a scanner and only watching win/loss, you're optimizing for the wrong metric and you won't know your model is broken until it's cost you real money. Happy to go deeper on any of these in the comments.