I can confirm this after hundreds of talks about the topic over the last 2 years. 90% of cases are simply not high-volume or high-stakes enough for the devs to care enough. I'm a founder of an evaluation automation startup, and our challenge is spotting teams right as their usage starts to grow and quality issues are about to escalate. Since that’s tough, we're trying to make the getting-to-first-evals so simple that teams can start building the mental models before things get out of hand.