Machine-checked mathematics: claims a computer verifies, not claims you take on trust. For a high-assurance reader this is the headline capability — the same discipline that formal methods bring to cryptographic protocols and safety-critical systems, applied to original mathematics.
How to read the status column. A result is PROVEN only when the project builds and
#print axioms <thm>shows the theorem reduces to Lean's foundational axioms (propext,Classical.choice,Quot.sound) with nosorryAx. A rawgrepfor the tokensorryis only an upper bound — it also catches the word in comments and docstrings, and in practice badly over-counts (the string "sorry-free" in a docstring reads as asorry). Every "proven" claim below was checked with#print axioms, not grep. Each certifiable repo carries aCertify.leanand a GitHub Actions workflow that rebuilds and re-runs the axiom check, failing onsorryAxor on a missing per-theorem report — public, reproducible proof a stranger can run, strictly stronger than "it ran on my laptop."
| Repo | src .lean |
theorems | status | note |
|---|---|---|---|---|
lean-sabidussi |
5 | 31 | PROVEN (0 sorry) |
PhD theorem. 4 headline theorems axiom-checked clean. The crown jewel. |
warnaar-glue |
7 | 88 | PROVEN (sorry-free) |
Glue for two Warnaar results; conditional theorems carry their literature inputs as explicit hypotheses, not axioms. |
lean-hopf-algebras |
15 | 157 | PARTIAL (honest) | 5 headline theorems axiom-clean; 3 sorry isolated in Transfer.lean, which the certified root does not build. |
social-network |
5 | 37 | PROVEN (sorry-free) |
Directed-container comonad laws (Haskell + Lean), no Mathlib dependency. |
trajectory-tracing |
25 | 227 | PROVEN (sorry-free) |
Epistemic-honesty formalism: 5 honesty/trajectory theorems, no Mathlib. |
trust-boundaries |
30 | 249 | PROVEN (sorry-free) |
Per-agent trust-level algebra (meet/join + cmax aggregation) — zero axioms. |
continued-fractions |
5 | 418 | STATED (not built) | Legendre palindrome / √n. Imports Mathlib but ships no Lake project — honestly left unbuilt rather than claimed. |
macbeth-seed |
85 | 631 | MIXED (44 sorry) |
Container-theory spine. A large fraction is scaffolding stated, not proven — said so plainly. |
symmetric-graphs |
42 | 365 | MIXED | Computational companion to lean-sabidussi: per-graph Hecke algebras + example catalogue. |
lean-sabidussiis the crown jewel. A non-trivial original theorem — the Sabidussi representation and Lorimer's theorem for vertex- and arc-transitive graphs, from my PhD — machine-checked, 0sorry. The four headline theorems (sabidussiIso,lorimer_forward,connectionSet_eq_doubleCoset,quotient_cosetGraph_iso) each reduce to[propext, Classical.choice, Quot.sound]and nothing else. The connection set the construction builds is a union of double cosets — a basis of the Hecke algebra ℋ(G,H) — so this repo is the formal core of the symmetric-graphs ⟷ Hecke-algebra link described on the landing page.macbeth-seedis stated honestly. 44sorryacross 631 theorems means a large part is a stated scaffold, not a proof. Calling all 631 "formalized" would be exactly the overclaim this index exists to prevent; the honest line is "a container spine with N proven, M stated."- The trust/honesty strand (
trust-boundaries,trajectory-tracing,social-network) is self-contained — no Mathlib, ~25-second CI builds — and fullysorry-free. These are the AI-trust results, formalized: that an honest trajectory keeps a non-degenerate trust level, that a circular trajectory collapses to ⊥, that trust composition is monotone.
Beyond the standalone repos above, the same mathematics is also formalized in Mathlib's own module structure and conventions, as a single local Mathlib development. This is kept local by deliberate choice — a personal development, not a set of public pull requests:
- Hopf algebras & umbral calculus — the additive (𝔾ₐ) Hopf/bialgebra structure on
R[X], delta operators and Rota's classification, umbral operators via generating functions, binomial-type sequences (ascending/descending Pochhammer, Lah), and the compositional inverse of a formal power series (Lagrange inversion). - Symmetric graphs / Sabidussi — group actions on simple graphs, coset graphs, the Sabidussi representation theorem, Lorimer's theorem, and quotient graphs.
- Named-graph & surface examples — the Langer graph and Tutte 12-cage; dodecahedron, Petersen, Heawood, Möbius–Kantor and other named graphs; voltage-graph constructions; the Meinhold and Zhou families as Sabidussi coset graphs; cellular surfaces and CSS quantum codes.
Mathlib-conformant in structure and naming, but not submitted upstream.
For Mathlib projects, pin a known-good lake-manifest.json to the target Mathlib
revision and lake exe cache get rather than running lake update from scratch; for
self-contained projects, just build (fast, no cache). Add a Certify.lean of
#print axioms statements, and a CI step that fails on sorryAx and on any missing
per-theorem report — accepting both depends on axioms: [...] and does not depend on any axioms (a zero-axiom theorem prints the latter). The per-theorem-report check
defeats a vacuous pass from a mistyped theorem name.
See also: the landing page for how this connects to the cryptography and trustworthy-ML work, and the expository index for the prose behind the proofs.