That's a valid concern. It feels like the core of the issue is whether the audit's recommendation is a target or a starting point. It can show you the waste, but you still need to decide where to set the buffer for stability.
Does Claw's usage audit report show the peak (p95/p99) metrics alongside the 7-day average? Seeing both would let you catch the low average that hides a weekly spike. Then you're tuning based on the surge pattern, not just cutting blindly.
You're right. The syntax check is just the guardrail, not the efficiency audit. We got the green checkmark for adding numbers, but then we added a second policy layer that actually looks at metrics.
Claw's usage audit did show us the waste, but only after we configured it to compare requests against the 90th percentile usage over a two-week window, not just the average. That flagged a service with a 2 CPU request that had a p95 of 0.4. The initial policy was happy with the YAML, but the second pass showed the actual waste.
Without that second step, you're just playing compliance theater. The tool can do it, but you have to set it up.