We just hit 10,000 scans. Here are the 5 biggest surprises from the data.

So we just crossed 10K scans and I've been digging through the dataset. What's the n? 10,000 data points is finally enough to stop hand-waving about patterns. Here's what jumped out at me: First — and this one's going to be controversial — our accuracy floor is way higher than we publicly claim. We're sitting at 94.7% precision on the core use case, but our marketing materials say "90%+". That's not a rounding error; that's leaving performance on the table. Second, latency variance is *brutal*. P95 is 340ms but P99 hits 2.8 seconds. That tail matters if we're pitching this for real-time applications. @Maya Chen, your team's optimization work bought us maybe 12% improvement, but we're still I/O bound. Third — and this surprises me — our false positive rate actually *improves* with batch size. Counter-intuitive, but the data's clean. We get 3.2% FP at batch size 1, drops to 1.1% at batch 32. That's a system design question, not a model question. Fourth, demographic variance is real and we need to talk about it. Our performance on the 18-35 cohort is 97.1%; on 55+ it's 91.3%. That's a 5.8-point spread that nobody's flagging. And fifth — regional differences are negligible (EU vs NA vs APAC within 1.2%), but *dataset source* matters enormously. Scans from mobile devices underperform by 4.7 points versus desktop. That's not a surprise if you think about lighting and angles, but it should change how we position this. Here's what I want to challenge: are we optimizing for the metrics that actually matter to customers, or the metrics that look good in a dashboard? @Frida Moreau, I know you've been talking to enterprise clients — what's their real constraint? Latency ceiling? Accuracy floor? Cost per scan? Because my reading of this data says we're solving for average case when someone's always going to care about worst case. What am I missing?