Logical independence
For any recursively enumerable and arithmetically sound axiomatic theory T, there exists an integer such that T cannot prove the values of BB(n) for any . Let denote the least such integer. For Zermelo–Fraenkel set theory (ZF), this value is known to be in the range:
This lower bound comes from the fact that BB(5) has been proven in Rocq.[footnote 1] The upper bound comes from an explicit TM which halts if and only if ZFC has a finite proof of contradiction. Assuming ZF is consistent and sound, then it cannot prove whether or not it is consistent, hence it cannot prove whether or not this specific TM halts.
Harvey Friedman spoke of embedding consistency statements within turing machines in a 2004 posting on the "Foundations of Mathematics" mailing list. Scott Aaronson conjectured in his Busy Beaver Frontier survey that and , where refers to the theory of Peano Arithmetic. Scott also mentioned (crediting Oscar Cunningham for the observation) that if there is an n-state machine that halts if and only if T is inconsistent, then + “BB(n) = b” can prove the consistency of T, where b is the true value of BB(n).[1]
Results by Theory
Most normal theories form a linear ordering with respect to consistency strength. All theories in the table below follow that linear order from top to bottom. Thus for any row X below Y, . Rows are empty if no better results is known than the result inherited from a previous row.
Note: For extremely weak theories, is not well defined without defining a specific TM encoding. Technically, for any theory requires fixing some TM encoding into the language of the theory. But for EFA and stronger theories, they are strong enough to prove the equivalence between all common TM encodings (ex: binary, beta-function, prime-power, MRDP), so specifying the precise encoding is generally irrelevant. Below EFA, the prime-power encoding cannot be decoded due to exponentiation not being provably total.
| Language | Theory | Upper bound on (states) |
|---|---|---|
| Set Theory | ZFC+I3 | |
| ZF+Berkeley cardinal | 1399 | |
| ZF+subtle cardinal | 493 | |
| Zermelo-Fraenkel set theory (ZF) / ZFC | 207 | |
| New Foundations (NF) + Axiom of Counting (AxCount) | 192[1] | |
| Zermelo set theory | 207 (via ZF)[footnote 2]
192 (via NF+AxCount)[footnote 3] | |
| New Foundations (NF) / Mac Lane | 185 | |
| Arithmetic | Peano arithmetic (PA) | 196 (direct) 185 (via NF)[footnote 4] |
| PRA / / / | ||
| EFA | ||
| / | ||
| / | ||
| / | ||
Axiom of Choice
Due to Shoenfield's absoluteness theorem, it is known that any TM proven non-halting in ZFC can also be proven non-halting in ZF (and the converse is trivially true), therefore
Therefore we refer to ZF and throughout this article since adding the Axiom of Choice does not have any effect on Turing machine decidability.
History
In a broad sense, the concept of a mechanical or computable procedure that was equivalent to the consistency of a formal theory was discovered by Kurt Gödel's incompleteness theorem, which is notable for introducing a meta-mathematical encoding of the statement that "T can prove statement S" (the provability predicate) which is used in all these machines. This was further clarified by Alan Turing and Alonzo Church's work on a formal definition of computability.
Although not a traditional turing machine, in 2012 Merlin Carl and Boris Z. Moroz constructed a Diophantine Equation which had solutions if and only if NBG (Von Neumann–Bernays–Gödel set theory) is inconsistent. NBG is a conservative extension of ZFC, so solving this equation is independent of ZFC.[2]
Zermelo–Fraenkel set theory (ZF)
There is no one authoritative source on the history of TMs independent of ZF, this is our best understanding of the history of TMs found. The results before 2026 are taken from Scott Aaronson's blog announcements and Busy Beaver Frontier or self-reported by the individuals who discovered them. The Aaronson-Yedida machine used a compiler called Laconic, which was then updated with NQL or "Not Quite Laconic", while the 432 machines use a compiler built by Andrew J. Wade. Many machines after 2026 have been generated by AI without a formal compiler.
Note that only the results with a linked formal verification are known to have such a verification, with the 7910 machine in particular relying on a result from Harvey Friedman that has no published proof.
| States | Date | Discoverer | Source | Verification |
|---|---|---|---|---|
| 7910 | May 2016 | Adam Yedidia and Scott Aaronson | Yedidia and Aaronson 2016[3] | |
| 1919 | 15 May 2016 | Stefan O'Rear | Github NQL file, Metamath Group Announcement | |
| 748 | May 2016 | Stefan O’Rear | Github NQL, Busy Beaver Frontier[1] | |
| 745 | July 2023 | Johannes Riebel | Riebel 2023 Bachelor Thesis[4] | |
| 643 | July 2024 | Rohan Ridenour | Github NQL, Aaronson Announcement | |
| 636 | 31 August 2024 | Rohan Ridenour | Github NQL (commit) | |
| 588 | 12 July 2025 | Andrew J. Wade | Github NQL (commit) | |
| 549 | 16 July 2025 | Andrew J. Wade | Github NQL (commit) | |
| 432 | 19 Aug 2025 | Andrew J. Wade | git commit | |
| 335 | 11 Oct 2026 | Jeff Ketchersid | discord message | |
| 254 | 13 Oct 2026 | Mira with use of ChatGPT, formalized by Jeff Ketchersid with Claude | Github | Rocq file |
| 230 | 19 Oct 2026 | Elliot Glazer with use of ChatGPT, formalized with GPT and Claude | Github | Rocq file |
| 215 | 21 Oct 2026 | Vladimir Reshetnikov formalized with Claude | Github | Rocq file |
| 209 | 12 Nov 2026 | James Chang formalized with ChatGPT | Github(discord message) | Lean |
| 211 | 17 Nov 2026 | Shivam Patel | Paper | Lean proof and replay |
| 207 | 19 Nov 2026 | James Chang formalized with ChatGPT | Lean proof and replay |

Wade's 432-state champion also omits the axiom of regularity, as ZF-regularity is also equiconsistent with ZF.
New Foundations (NF)
New Foundations is a formal theory created by W.V.O. Quine in 1937, based around the principle of stratified comprehension of sets. It is known to be stronger than Peano Arithmetic and second-order arithmetic, but weaker than ZF. The Github page contains a proof in Lean that it is equiconsistent with Mac Lane set theory, which is a novel result as of the time of writing (November 2026). Mac Lane set theory is far more mainstream than NF, so the following machine could be seen as an efficient way of determining which Busy Beaver value MAC cannot verify.
| States | Date | Discoverer | Source | Verification |
|---|---|---|---|---|
| 210 | 8 November 2026 | @wlad-svennik | [2] | |
| 198 | 12 November 2026 | @wlad-svennik | [3] | |
| 192 | 16 November 2026 | @wlad-svennik | [4] | |
| 185 | 19 November 2026 | @wlad-svennik | Github | Lean verification on Github (currently a 20MB tarball) |
User-friendly theory
The TM searches for an inconsistency in the theory "FunctionNF". It is defined to equal stratified λ-calculus minus pairing, minus case-expressions, plus full functional extensionality, plus an equality checker.
Terms t ::= x | t u | λx.t | (s ≐ t) | ⊤ | ⊥
Depth Each subterm occurrence has a depth relative to the term it sits in:
the term itself 0
in t u at depth n: t at n+1, u at n
in λx.t at depth n: t at n−1
in s ≐ t at depth n: s and t at n
Formation λx.t is a term iff every free occurrence of x in t has depth 0 in t.
Nothing else is constrained: other variables may occur at any depths.
Axioms (β) (λx.t) u = t[u/x]
(ext) (∀x. f x = g x) → f = g
(≐) s = t → (s ≐ t) = ⊤ s ≠ t → (s ≐ t) = ⊥
(nt) ⊤ ≠ ⊥
The logic is ordinary first-order logic with equality, so unstratified formulas are allowed. Only λ is restricted.
It's not obvious, but the above suffices to prove the Axiom of Infinity, i.e. that the set of natural numbers exists and has all of its expected properties. It's then possible to define the set of real numbers in any of the usual ways, and do nearly all of extant mathematics.
NF vs Mac Lane vs other set theories
The Github link contains proofs in Lean that:
- NF interprets FunctionNF (which is very short),
- FunctionNF in turn interprets NFU+Infinity (which is similar to Specker's proof of the "Axiom of Infinity" within NF),
- NFU+Infinity interprets NF, which is a novel and hard result that is not yet found in the literature,
- NFU+Infinity interprets Mac Lane set theory,
- Mac Lane set theory interprets TSTI (Theory of Simple Types plus (Axiom of) Infinity),
- TSTI is equiconsistent with NFU+Infinity.
So overall, the machine halts iff there exists an inconsistency in any of the above theories. Regardless, Mac Lane set theory is actually weaker than NF in terms of intepretability strength, as opposed to just consistency strength.
With Rosser's Axiom of Counting
If we add the Axiom of Counting (AxCount) to NF, which is the rather tautological-looking axiom
For any finite set S, |S| = |{{x} : x ∈ S}| (i.e. the number of elements of S doesn't change when you replace each element x with the singleton {x})
then it's sufficient to prove Con(Z). As of the time of this writing (November 2026) we've used this to obtain a machine that is smaller (192 states) than any machine of Zermelo-strength or above.
ZFC can prove Con(NF + AxCount).
Implementation theory: The combinatory calculus
FunctionalNF is not implemented directly. Instead, a massively leaner form is implemented, in the form of a combinatory logic which consists of just three nonlogical primitives:
- Binary application , associated to the left.
- A constant . This is a fixed-point tester.
- A binary function . This is a merger between the K combinator and a Skolem function for function extensionality.
Define the convenient shorthands
The six axioms, with all free variables universally quantified, are:
| No. | Axiom | Role |
|---|---|---|
| 1 | Constant functions | |
| 2 | Combinatory composition | |
| 3 | Positive equality test | |
| 4 | Negative equality test | |
| 5 | Extensionality witness | |
| 6 | Nondegeneracy |
In particular, follows immediately from axiom 1. It is not a separate axiom.
With Rosser's Axiom of Counting
The 196-state machine uses the same three primitives. In the six-axiom calculus above, serves both as ⊤ and is meant to satisfy axiom 2. By contrast, the Counting variant uses for ⊤, and makes honour axiom 2 by defining . The definitions and carry over.[5]
With this definition of , axioms 1, 2, 4 and 5 carry over as written. Replacing axioms 3 and 6 and adding axiom 7 gives the following changes. All free variables are universally quantified.
| No. | Axiom | Role |
|---|---|---|
| 3 | Positive fixed-point test | |
| 6 | Nondegeneracy. The machine halts if it derives . | |
| 7 | Counting |
Thus has two roles. It serves as ⊥, and it is the operator in the Counting axiom. To read axiom 7, put . The axiom then says that for every in the range of , applying to the fixed argument recovers .
Fragment of First-Order Logic
Notably, we do not implement the whole of First-Order Logic: The quantifier ∀ can be defined internally to CombNF / FunctionalNF:
∀ := λP.(P ≐ λx.⊤)
¬ := λp.(p ≐ ⊥)
∃ := λP. ¬ (∀ λx. ¬ (P x))
Moreover, we've avoided implementing variables thanks to our use of combinatory logic. The only fragment of First-Order Logic we do implement consists of "=", the above function symbols, and propositional logic. This is similar to Primitive Recursive Arithmetic and System T.
What the symbols do
is an equality tester:
has two jobs. On the diagonal it constructs constant functions:
Off the diagonal it supplies a witness to distinguish two functions whenever they are different:
supplies the composition law
Here and are terms, not primitive constants.
Once again, we do not implement the whole of First-Order Logic.
The formalized proof for this machine's NF-undecidability is formalized as an external Lean metatheorem rather than a completed PA-internal derivation. As of this publication, a fresh independent clean build and community review have not been reported.
Peano Arithmetic (PA)
The ZF bound applies to PA because ZF proves the arithmetical theorems of PA. In turn, a proof of an exact Busy Beaver value in any theory listed below PA would give a PA proof of that same value. For , a proof of would let PA decide whether the 196-state machine halts by padding it to states and checking its first steps.
| States | Date | Discoverer | Source | Verification |
|---|---|---|---|---|
| 372 | 11 February 2026 | @LegionMammal978 (Matthew House) | Github | |
| 196 | 12 Nov 2026 | James Chang formalized with ChatGPT | Github | Lean |
This machine takes Wade's ZF champion and replaces the set theory axioms with one axiom schema of "adjunction + separation", which has been shown to interpret PA.[6]
The 185-state machine halts exactly when NF proves a contradiction.[7] The Holmes–Wilshaw consistency proof for NF therefore establishes in ZFC that the machine runs forever.[8] For this same machine, PA proves that if it runs forever, then PA is consistent.[9] By Gödel's second incompleteness theorem, PA cannot prove that it runs forever. If PA proved a numerical upper bound on , it could verify that the machine survives that many steps and hence prove that it runs forever. Thus .
Weaker arithmetical theories
Even , a proper subtheory of Tarski–Mostowski–Robinson , proves every true individual halting assertion because it proves all true sentences in the usual language of arithmetic.[10] Between Robinson arithmetic and bounded induction lie , the theory of discretely ordered semirings, and , which adds induction for quantifier-free formulas.[11] Further up, adds total superexponentiation to EFA. It proves the consistency of , which EFA cannot prove.[12]
Theories grouped in one arithmetic row have the same Busy Beaver threshold. For fixed standard and , a counterexample to is a halting computation of an -state machine lasting more than steps. A proposed computation can be checked in polynomial time in the length of its complete history. When is true, a particular halting run witnesses the matching lower bound.
With this explicit history coding, and agree on the universal equations expressing the upper bounds. Full bounded arithmetic and agree on sentences in their common arithmetical language.[13] Buss's conservation theorem similarly gives and the same provable upper bounds.[14] Finally, , and agree with PRA on arithmetic sentences, which suffices for all four to share a threshold.[15]
Large cardinals
These tables are for theories stronger than ZFC.
| States | Date | Discoverer | Source | Verification |
|---|---|---|---|---|
| 493 | 25 February 2026 | @LegionMammal978 | Github |
This machine takes Wade's ZF champion and adds the axiom of Choice and a statement from Harvey Friedman (Proposition 4.3 of "Primitive Independence Results")[16] which Friedman showed to be equivalent to "There exist arbitrarily large subtle cardinals" over ZFC.
For reference, this theory is between the strength of "strongly unfoldable" and " exists" on the large cardinal hierarchy on Wikipedia.
Note: The initial machine produced in 2016 by Yedidia and Aaronson is designed to halt iff a certain statement created by Harvey Friedman is true.[1] According to Friedman (without a published proof), this statement is independent of the theory "Stationary Ramsey Property" or SRP, which is equiconsistent with the theory which is also between the strength of "strongly unfoldable" and " exists" on the large cardinal hierarchy.[17] However, the theory used by the BB(493) machine is also independent of this theory.
| States | Date | Discoverer | Source | Verification |
|---|---|---|---|---|
| 1399 | 8 November 2026 | @Meganeuridae | Github | Rocq file |
Berkeley cardinals are perhaps the strongest family of large cardinals considered, initially proposed by Hugh Woodin at a lecture at UC Berkeley as an axiom that was "obviously inconsistent".[citation needed] However, no inconsistency has been found and many people now believe it may be consistent. Berkeley cardinals are known to be inconsistent with the axiom of choice.
Footnotes
- ↑ The fact that this theorem has been proven in Rocq technically does not imply that the theorem must be provable in ZF, because the consistency strength of Rocq is actually higher than that of ZF. However, the BB(5) proof does not use any techniques that could not be formalized within ZF.
- ↑ Zermelo set theory is a subtheory of ZF. For the version with Choice, the same upper bound follows from inclusion in ZFC and the equality of the ZF and ZFC thresholds.
- ↑ The 192-state NF + AxCount machine also gives for Zermelo set theory. Writing for its blank-tape halting statement, the note proves . Thus, if is consistent, it cannot prove that the machine runs forever. The note derives actual nonhalting in ZF using the cited consistency comparisons. Proof and verification files
- ↑ The 185-state bound also applies to every weaker arithmetic theory listed below PA, since any numerical upper bound on provable in one of those theories is also provable in PA. See the PA section for the proof of the 185-state bound.
References
- ↑ 1.0 1.1 1.2 Scott Aaronson. 2020. The Busy Beaver Frontier. SIGACT News 51, 3 (August 2020), 32–54. https://doi.org/10.1145/3427361.3427369
- ↑ Carl, M., & Moroz, B. Z. (2012). On a Diophantine representation of the predicate of provability. Zapiski Nauchnykh Seminarov POMI, 407, 77-104. (Republished in 2014 in the Journal of Mathematical Sciences) (preprint)
- ↑ A. Yedidia and S. Aaronson. A relatively small Turing machine whose behavior is independent of set theory. Complex Systems, (25):4, 2016. https://arxiv.org/abs/1605.04343
- ↑ Riebel, Johannes (March 2023). The Undecidability of BB(748): Understanding Gödel's Incompleteness Theorems (PDF) (Bachelor's thesis). University of Augsburg.
- ↑ A 201-state Busy Beaver bound for Zermelo set theory, proof and verification files.
- ↑ https://mathoverflow.net/questions/508137/interpreting-pa-with-only-adjunction-separation/508161
- ↑ https://github.com/wlad-svennik/notes/tree/bfc960b0725ecab9be07518cc35954662f2592e7/nf-bb185/pa
- ↑ M. Randall Holmes and Sky Wilshaw, NF is Consistent.
- ↑ https://github.com/wlad-svennik/notes/tree/bfc960b0725ecab9be07518cc35954662f2592e7/nf-bb185/pa
- ↑ Taishi Kurahashi and Albert Visser, Certified Σ₁-Sentences, §2, Theorem 2.3.
- ↑ Kazuyuki Tanaka, Logic and Computation I, Part 3a. Formal Arithmetic, Corollary 3.43.
- ↑ Samuel R. Buss, First-Order Proof Theory of Arithmetic, §4.3.
- ↑ Jan Krajíček, Bounded Arithmetic, Propositional Logic, and Complexity Theory, Lemma 5.2.15 and Theorem 5.3.4.
- ↑ Samuel R. Buss, Axiomatizations and Conservation Results for Fragments of Bounded Arithmetic.
- ↑ Jeremy Avigad, Notes on Π¹₁-conservativity, ω-submodels, and the collection schema, §2.
- ↑ H. Friedman, "Primitive Independence Results" (2002). Accessed February 2026.
- ↑ tlonuqbar, "What is the consistency strength of SRP (Stationary Ramsey Property)?" (2026). MathOverflow post, accessed February 2026.