cat _posts/2026-10-15-leanfrontier-field-note-25-and-the-kernel-audit-becomes-a-job.md
LeanFrontier: Field Note 25, and the kernel audit becomes a job
LeanFrontier’s Field Note 25, Fixed Before It Broke., tells the v4.35.0 upgrade as one dry run: the audit against v4.35.0-rc4 on 9 October failed on a single line of ChungErdos (MemLp.indicator now takes a NullMeasurableSet), a two-version proof was written a week early so the release could land unattended once Lean v4.35.0 and Mathlib appeared on 12–13 October, and #503 then kept the v4.35 proof alone (#508, fe01993).
The measurements taken for issue #504 are now a job rather than a note: .github/workflows/kernel-audit.yml replays the corpus, with its whole Mathlib closure, through Lean’s kernel with lake check — monthly, and as of #507/95f14c6 once after each upgrade lands, on a push to main touching lean-toolchain, deliberately outside the blocking upgrade check it would more than double (a3dfed7). It runs the default kernel only (#506, a9125d3): the first run on an ubuntu-26.04 runner took two hours and 10.3 GB, while --paranoid took 3 h 19 min on a laptop and would exceed the six-hour job limit, so the bundled independent checkers are run by hand — and nanoda, one of them, dies with a stack overflow rather than a verdict while the other five accept the whole export. The job audits already-admitted code, so it changes no verdict and blocks nothing.
The same day produced the releases. v0.3.0 is 110 accepted modules on Lean and Mathlib v4.35.0, up from 102 in v0.2.0, and the matching data release corpus-v2 holds 110 submissions, 110 modules and 90 internal import edges — with the caveat its own note insists on, that every record predates 7 October and so carries triviality-probe outcomes from the old probe, which stated each goal without its module’s namespace and context; re-reading all accepted entrypoints in their real context found no submission the corrected probe would reject.