What does a compiled Lean proof actually certify?

With the recent Riemann hypothesis claim backed by Lean code, I keep thinking about this. I'm not saying the claim is wrong, I'm asking what the check actually covers.

Lean checks that each step follows. It doesn't check that the statement and definitions mean what we think they mean. And some conventions make that riskier in Lean, 1/0 = 0 and 3 − 5 = 0 in ℕ. A definition that quietly hits one of those junk values can compile perfectly and still prove something vacuous or different.

Also, a 100k-line Lean proof tells you that something is true, but on its own it doesn't tell you why. I'd still take a clean 5-page hand proof over that any day.

So my questions:

  1. Who actually checks that the formal statement matches the informal one?

  2. Should a Lean proof count as accepted without a human-readable proof?

Curious what people here think that.

Edit: Andrew Sutherland (MIT) has posted a live status page for checking the Lean proof: https://math.mit.edu/~drew/grh/ It answers most of the questions raised in this thread. Per the page all 175,187 modules compile, the main theorems depend only on the three standard axioms (no native_decide), the statement is Mathlibs own RiemannHypothesis rather than a redefinition the Mathlib files match the official cache byte for byte, and the independent Rust kernel nanoda accepts the full export. Some kernel checks are still running, and the page itself notes that a pass means follows from the standard axioms according to the kernels used, still relying on Mathlib's definitions meaning what we intend. Also, GRH there means Dirichlet L-functions only. So on the formal side, this is close to the best case one could ask for. What remains open is the part a kernel can't give us a readable, structural proof that shows why it's true. Thanks to everyone for the great discussion 🫶✨️

Author: DifficultAvocado6185