Your repo is a preference dataset: extracting taste from merge history

You're spending less time thinking 'Can we build this?' but more asking 'Which of all the possibilities should we build?' Now taste bottlenecks execution. And eliciting preferences from experts is expensive but what if you could extract them from the versioned artifacts you've been maintaining all along? Under a mild structural assumption that your team's trajectory of accepted revisions is directionally improving in expectation, you can distill preferences into your agents. Implicit Preference Distillation facilitates cheaply aligning your AI with your institutional practices. We're experimenting with extracting preference signals from a repo merge history, but the same strategy applies anywhere you're iteratively refining artifacts toward a quality bar.