In 1998, Thomas Hales announced a proof of a claim Johannes Kepler made in 1611: no arrangement of equal spheres fills space more densely than the familiar pyramid of oranges on a grocer's stall. The proof combined hundreds of pages of argument with gigabytes of computer calculations.

A panel of twelve referees studied it for four years. Their verdict was that they were 99 percent certain it was correct. That is a strange number in mathematics. A theorem is either proved or not proved. Yet here the best human checkers could go no further, because the argument was simply too large to inspect completely.

This raises a serious question. If expert human checking has limits, what grounds mathematical certainty? One answer lies in proof assistants, programs that verify every logical step of an argument mechanically. Below, we examine how they encode mathematics, how the major systems differ, and which landmark results they have placed beyond reasonable doubt.

Formal Verification: Every Step Made Explicit

A proof on paper is written for a human reader. It skips steps the author judges obvious, relies on shared conventions, and trusts the reader to fill the gaps. A proof assistant grants no such trust. Every definition must be built from foundations, and every inference must be justified by an explicit rule. Only then can a machine confirm, mechanically, that the conclusion follows.

Consider the natural numbers as Peano defined them: zero is a number, and every number has a successor. Addition is then defined by two equations: n + 0 = n, and n + S(m) = S(n + m). Now, n + 0 = n holds immediately by definition. However, 0 + n = n does not. It requires a proof by induction on n. A human calls both facts obvious. The proof assistant distinguishes them precisely, and so it shows exactly where reasoning is required.

Many modern systems rest on an observation known as the Curry–Howard correspondence: propositions can be read as types, and proofs as programs belonging to those types. A proof of “A implies B” becomes a function that transforms evidence of A into evidence of B. As a result, checking a proof reduces to checking that a program has the type it claims. That task is mechanical, carried out step by step without any need for insight.

Writing every inference by hand would be unbearable, so users issue tactics. These are commands such as “apply induction” or “simplify” that generate the low-level steps automatically. Yet the tactics need not be trusted. Under a principle known as the de Bruijn criterion, everything is finally passed to a small kernel, often a few thousand lines of code, and the kernel alone decides whether a proof is accepted. Our confidence therefore rests not on the whole system but on one compact, heavily scrutinised component.

Takeaway

A formal proof moves trust away from the cleverness of the argument and onto the simplicity of the checker. You need not understand a proof's ingenuity to be certain it is valid.

Three Systems, Three Architectures

Proof assistants differ chiefly in their logical foundations, and those foundations shape everything built on top of them. Three systems illustrate the main design choices: Coq, Lean, and Isabelle.

Coq has been developed at INRIA in France since the 1980s and was recently renamed Rocq. It is built on the Calculus of Inductive Constructions, a rich type theory devised by Thierry Coquand and Gérard Huet. Its logic is constructive by default: to prove that something exists, one must in principle exhibit it. Classical principles, such as the law of excluded middle, can be added as axioms when desired. Because its proofs are programs, Coq can extract verified code. It underpins CompCert, a C compiler proved to translate programs faithfully.

Lean was created by Leonardo de Moura at Microsoft Research in 2013. It uses a closely related dependent type theory, but it has been embraced largely by working mathematicians. Its community library, mathlib, contains hundreds of thousands of formalised theorems across algebra, analysis, topology, and number theory. Lean users adopt classical reasoning freely. The system aims at the mathematics found in research papers and graduate textbooks rather than at constructive purity.

Isabelle was begun by Lawrence Paulson at Cambridge and developed further with Tobias Nipkow's group in Munich. It takes a different route. Its most widely used version, Isabelle/HOL, rests on higher-order logic, which is simpler than dependent type theory and therefore easier to automate. Its tool Sledgehammer sends subgoals to external automatic provers, then rebuilds their answers inside the trusted kernel. Isabelle was used to verify the seL4 operating system kernel, showing that formal methods scale to industrial software. In short, expressive power and automation pull against each other, and each system strikes its own balance.

Takeaway

The choice of logical foundation is not a technicality. It determines what can be said easily, what can be proved automatically, and what counts as a proof at all.

Landmark Verifications: From Doubt to Certainty

The four-colour theorem states that any map drawn in the plane can be coloured with four colours so that no two neighbouring regions share a colour. In 1976, Kenneth Appel and Wolfgang Haken proved it by reducing the problem to nearly two thousand configurations, each checked by computer. Many mathematicians objected. No human could inspect all the cases, and the checking programs might themselves contain errors. In 2005, Georges Gonthier and Benjamin Werner completed a full formal proof in Coq. It verified both the mathematical reasoning and the case analysis within a single trusted framework.

The Kepler conjecture followed a similar path. After the referees' “99 percent” verdict, Hales launched the Flyspeck project to remove the remaining doubt. Using the proof assistants HOL Light and Isabelle, an international team formalised the entire argument, including the computer calculations at its core. The project was completed in 2014, and the formal proof was published in 2017. A conjecture stated in 1611 thus reached a level of certainty that no panel of referees could have provided.

Other landmarks show formal methods moving into more conceptual mathematics. In 2012, Gonthier's team formalised the Feit–Thompson odd order theorem, a cornerstone of finite group theory whose original proof ran to about 255 pages. In 2020, Peter Scholze challenged the Lean community to verify a central theorem from his work with Dustin Clausen, a result he himself was not fully confident in. Within six months the crucial part had been checked, and the full verification was finished in 2022. The process even clarified parts of the original argument.

Nevertheless, one gap remains that no machine closes. A proof assistant certifies that a formal statement follows from the axioms. It cannot certify that the formal statement says what we intended it to say. Human judgement still governs the definitions, while the machine governs the inferences.

Takeaway

Machine verification answers “does this follow?” with near-perfect reliability. The question “is this what we meant?” remains a human responsibility.

Return to the referees' “99 percent.” Mathematics has always aimed at certainty, and Euclid showed the method: fix the axioms, state the rules, and let every conclusion follow by necessity. Proof assistants carry that method to its logical end by requiring every step to be justified explicitly.

The result is a division of labour. Humans supply insight, choose definitions, and design the structure of an argument. The machine checks each joint, without fatigue and without wishful thinking.

This certainty is not absolute in a philosophical sense, since it rests on chosen axioms and on a small kernel we must trust. Yet it is certainty of the most transparent kind: every brick placed, and every brick inspected.