一份一千七百万行的证明,没有人读得完;但谁都可以让机器替自己从头查一遍。本页把“怎么查”画成两个场景:CI 怎样把 Main 切成可以分头编译、凭内容复用的分片;读者怎样在自己的机器上,不信任 CI、不信任作者,把这些分片重放一遍。本页是示意模型加真实数字,不是证明。
git clone、./verify --status Main、可先试小模块 ./verify BuildAid.D1、再 ./verify Main。脚本从 16 个 Release 拉回分片,逐片核对 Release 公布的 SHA-256;再用 lake env leanchecker --fresh 在空白的 Lean 环境里把模块及其导入的每条声明重放一遍;最后核对五条顶层定理的公理只落在 propext、Classical.choice、Quot.sound 之内(ci/Audit.lean 用 #print axioms 做同样的事)。src/Challenge.lean 只依赖 Mathlib v4.33.1(提交 0df444a360ea),写下五条陈述;Comparator 核对 Main 证明的类型与它逐条相同,公理不出那三条,再交给 Lean 4.36 内核与 leanchecker-paranoid、nanoda、con-leche、con-ron 在断网的 bubblewrap 沙盒里各查一遍;lean4lean 用 Lean v4.33.1 另行构建、单独运行。这条多校验器路线在整个提交上的完整运行是数小时的长任务,本页只画它的规则。Nobody can read a proof of seventeen million lines, but anyone can have a machine check it from scratch for them. The page draws how in two scenes: how CI cuts Main into shards that compile separately and are reused by their content, and how a reader replays those shards on their own machine, trusting neither CI nor the author. The page is an illustrative model with real figures, not a proof.
git clone, ./verify --status Main, optionally the small module ./verify BuildAid.D1, then ./verify Main. The script pulls the shards from 16 Releases and checks each against the SHA-256 the Release publishes; lake env leanchecker --fresh then replays every declaration of the module and its imports in an empty Lean environment; last, it checks that the five headline theorems use no axioms beyond propext, Classical.choice and Quot.sound (ci/Audit.lean does the same with #print axioms).src/Challenge.lean depends on Mathlib v4.33.1 (commit 0df444a360ea) alone and writes down the five statements; Comparator checks that the types proved in Main are those statements, that no axiom beyond the three is used, and hands the proof to the Lean 4.36 kernel and to leanchecker-paranoid, nanoda, con-leche and con-ron in a bubblewrap sandbox without network; lean4lean is built with Lean v4.33.1 and run separately. A full run of this route on the whole submission takes many hours; the page draws its rules only.In the book: Chapter 7 · the 35-line challenge contract · 8,490 shards and honest boundaries
RiemannHypothesis 陈述:若 ζ(s) = 0,s 不是负偶数,且 s ≠ 1,则 Re(s) = 1/2。已由 Lean 内核验证:The Riemann Hypothesis, as Mathlib states it (RiemannHypothesis): if ζ(s) = 0, s is not a negative even integer and s ≠ 1, then Re(s) = 1/2. Verified by the Lean kernel: riemannHypothesisRiemannZetaZerosOnLine:对一切复数 s,若 ζ(s) = 0,s 不等于任何 −2(n + 1)(n 为自然数),且 s ≠ 1,则 Re(s) = 1/2。已由 Lean 内核验证:RiemannZetaZerosOnLine: for every complex s, if ζ(s) = 0, s is no −2(n + 1) with n a natural number, and s ≠ 1, then Re(s) = 1/2. Verified by the Lean kernel: riemannZeta_eq_zero_reDirichletLNonvanishingRightOfHalf:对一切 N ≠ 0、一切模 N 的狄利克雷特征 χ 与一切复数 s,若 1/2 < Re(s),且 χ ≠ 1 或 s ≠ 1,则 L(χ, s) ≠ 0。已由 Lean 内核验证:DirichletLNonvanishingRightOfHalf: for every N ≠ 0, every Dirichlet character χ modulo N and every complex s, if 1/2 < Re(s) and χ ≠ 1 or s ≠ 1, then L(χ, s) ≠ 0. Verified by the Lean kernel: LFunction_ne_zero_of_half_lt_reDirichletLZerosOnLine:对一切 N ≠ 0、一切模 N 的狄利克雷特征 χ 与一切复数 s,若 0 < Re(s) < 1 且 L(χ, s) = 0,则 Re(s) = 1/2。已由 Lean 内核验证:DirichletLZerosOnLine: for every N ≠ 0, every Dirichlet character χ modulo N and every complex s, if 0 < Re(s) < 1 and L(χ, s) = 0, then Re(s) = 1/2. Verified by the Lean kernel: LFunction_eq_zero_rePrimitiveDirichletLZerosOnLine:对一切 N ≠ 0 与一切模 N 的本原狄利克雷特征 χ,以及一切复数 s,若 0 < Re(s) < 1 且 L(χ, s) = 0,则 Re(s) = 1/2。已由 Lean 内核验证:PrimitiveDirichletLZerosOnLine: for every N ≠ 0, every primitive Dirichlet character χ modulo N and every complex s, if 0 < Re(s) < 1 and L(χ, s) = 0, then Re(s) = 1/2. Verified by the Lean kernel: LFunction_eq_zero_re_of_isPrimitive