How do you write meaningful performance assertions in your load tests?

Most load test examples I see just report numbers (p95, p99, RPS) without failing the test definitively. I want our performance tests to actually gate CI. Current approach: k6 thresholds like `p(95) < 500`. But: - How do you define the threshold? We don't have a formal SLA. - How do you handle variance — an external network blip shouldn't fail the build. - Should performance tests block PRs or just run in a nightly pipeline? Looking for the thought process, not just the syntax.