A pipeline that takes a theorem written in prose, formalizes it into
Lean 4, compiles it against the kernel, and then audits which axioms the result
actually depends on. Below are recorded runs, replayed exactly as
they happened — including the ones that failed.
Recorded runs
73
Verified
44
Rejected
6
Kernel
Lean 4
Library
Mathlib
replay of recorded runs — nothing is executed in your browser
Mechanism
What the checker actually does
The claim is not "an assistant thinks this proof is right". It is that a
kernel accepted a formal statement, and that an audit enumerated every axiom the
accepted proof rests on. Each stage can reject.
The guard is the step that makes the rest meaningful: a
formalization that quietly weakens the statement would compile perfectly and prove
nothing. Rejection at any stage returns a diagnosis rather than a verdict — the
fourth run below is a rejection whose diagnosis found a bug in the harness, not in
the mathematics.
Recorded runs
Four runs, start to finish
Pick a run to see what went in, the Lean that came out, and what the
kernel and the audit said about it.