Benchmark route
Every table cell is a mean over images. These are the distributions those means come from, so you can see whether a number describes a typical image or is an artefact of a lopsided tail.
Each row is one model on one route: a histogram of its per-image score in 20 bins, with the mean marked by a solid line and the median by a dotted one. The gap between them is the whole point. A mean below the median means a minority of bad images is dragging the average down and the model is typically better than its headline number. A mean above the median means the opposite: most images are worse than the average, which is being propped up by a minority that work.
Detecture's mean understates it on every route. Mean minus median is negative everywhere — −0.086 on CSTD, where the median image scores 0.969 against a mean of 0.884. A small tail of genuinely bad images pulls the average down; the typical image is markedly better than the table suggests.
Sa2VA is the reverse, and severely so. On RWTD-COCO its mean is 0.240 but its median is exactly 0.000: it returns nothing at all on 63% of images, and the aggregate is carried entirely by the minority it does answer. Its score is not a typical performance, it is an average over a bimodal split. The same holds on TextureADE, where 42% score zero.
LISA's spread is the tightest of any model, and that is the finding. On RWTD-COCO its p10 is 0.442 and its p90 is 0.503 — almost every image lands in a band 0.06 wide. That is not robustness, it is the signature of returning one mask of roughly constant area regardless of the image. Nothing is biasing LISA; it is uniformly the same answer.