cat _posts/2026-10-15-leanfrontier-the-corpus-moves-to-mathlib-v4-35-0-and-the-threat-model-names-what-stays-unguarded.md
LeanFrontier: the corpus moves to Mathlib v4.35.0, and the threat model names what stays unguarded
LeanFrontier is on Lean/Mathlib v4.35.0 (#502, fdba09a). The weekly upgrade merged unattended, as the v4.34 one did in September: the required mathlib-upgrade check rebuilt the fingerprint index and blocked any newly exact duplicate in the accepted corpus, the audit accepted all 305 entrypoints with no collisions and no corpus deprecations, and the kernel recheck and downstream smoke passed. The follow-up is the separation of statement from proof the corpus insists on — ChungErdos keeps only the v4.35 proof, since the v4.34 alternative added for compatibility in #500 can no longer elaborate on main and had drawn an unused-tactic lint (#503, 4a6eaa2). That the change is proof-only is evidence rather than assertion: lake exe frontier-audit reports the same 1,184 declarations with identical canonical types, kinds and axioms before and after, and leanchecker passes on the module.
Two entries then landed in docs/threat-model.md under what remains unguarded (#505, 5bb3d65). One kernel: admission still uses only leanchecker, and the independent kernels that ship with Lean 4.35 arrive as lake check, which exports the targets and replays them — but it replays the whole Mathlib closure, and measuring that against the corpus costs about an hour and 10.7 GB of peak memory for lake check and 3 h 19 min and 12.2 GB under --paranoid. That makes them a periodic corpus audit rather than an admission step, and no stronger claim is made until a job actually runs them. The measurement is also where the interesting failure is: five of the six paranoid checkers — Lean’s paranoid kernel, lean4lean, con-leche, con-ron and the default kernel — accept the whole export, and nanoda dies with a stack overflow before reaching a verdict, which fails the run without being a rejection. Source syntax: the statement reader depends on today’s file syntax, Lean’s module system is opt-in at 4.35 and every corpus file is already warned that Mathlib expects it, so adopting it would need the reader and the weekly statement screen to learn public import and visibility first.