Verifier Theory and Unverifiability

Despite significant developments in Proof Theory, surprisingly little attention has been devoted to the concept of proof verifier. In particular, the mathematical community may be interested in studying different types of proof verifiers (people, programs, oracles, communities, superintelligences) as mathematical objects. Such an effort could reveal their properties, their powers and limitations (particularly in human mathematicians), minimum and maximum complexity, as well as self-verification and self-reference issues. We propose an initial classification system for verifiers and provide some rudimentary analysis of solved and open problems in this important domain. Our main contribution is a formal introduction of the notion of unverifiability, for which the paper could serve as a general citation in domains of theorem proving, as well as software and AI verification.

💡 Research Summary

The paper opens by observing that modern proof theory has largely ignored the entity that actually checks proofs – the verifier – and argues that verifiers themselves deserve systematic study as mathematical objects. It proposes a functional model of a verifier as a mapping from a proof (the input) to a verification result (the output) and then classifies verifiers according to the substrate that implements them. Five broad categories are identified: (1) human mathematicians, (2) automated theorem provers (ATPs), (3) external oracles, (4) scholarly communities or peer‑review collectives, and (5) hypothetical super‑intelligent AIs. For each category the authors introduce four analytical dimensions: expressive power (what proof languages the verifier can understand), computational complexity (time and space required for verification), reliability (error rates and typical failure modes), and self‑verification capability (whether the verifier can certify its own verification process).

The discussion of humans emphasizes cognitive limits: bounded working memory, susceptibility to confirmation bias, and a practical time‑complexity ceiling that restricts reliable verification to proofs that can be processed within polynomial‑time mental effort. Automated provers are examined through the lens of complexity theory; depending on the underlying logic they may operate in PSPACE, EXPTIME, or higher, but they remain vulnerable to implementation bugs and to the incompleteness of the formal system they are built upon. Oracles are treated as idealised information sources that could, in principle, answer any decidable query, yet any concrete realization will inevitably be incomplete or erroneous, making the oracle’s trustworthiness a critical factor. Communities are modelled as networks of human and machine verifiers that exchange verification judgments; while collective intelligence can raise overall confidence, phenomena such as echo chambers, majority‑vote errors, and coordination failures can degrade reliability. Super‑intelligent AIs are presented as a theoretical upper bound: they could, in principle, verify any proof within a given formal system, but Gödel’s incompleteness theorem and self‑reference paradoxes impose a fundamental barrier to full self‑verification, leading to an infinite regress problem.

A central contribution of the paper is the formal definition of “unverifiability.” Unverifiability is the property of a proof (or class of proofs) for which no verifier in a specified set can produce a correct verification result. Two illustrative cases are explored. First, human verifiers cannot reliably check proofs that exceed their cognitive or temporal limits, such as extremely long computer‑assisted proofs (e.g., the four‑color theorem) where the sheer volume of low‑level steps outstrips human capacity. Second, a program attempting to verify its own verification algorithm encounters a Gödel‑type limitation: any sufficiently expressive system that can encode its own correctness proof will inevitably contain statements that are true but unprovable within that system, rendering full self‑verification impossible.

The authors then introduce a “verification network” model in which multiple verifiers interact, exchange results, and reach consensus through mechanisms such as majority voting, weighted trust scores, or cryptographic protocols (e.g., blockchain‑based zero‑knowledge proofs). They prove that a network can achieve maximal practical reliability only if it contains at least one verifier that is effectively omniscient for the target proof class—a condition that is unattainable in reality. Consequently, verification networks must be designed to approximate the best possible reliability given the inherent unverifiability of certain proof fragments.

In the concluding section the paper outlines a research agenda. It calls for quantitative models of verifier complexity and reliability, experimental studies of human‑machine collaborative verification, formal investigations into ways to circumvent or mitigate self‑verification barriers, and the development of robust consensus protocols that balance security, privacy, and scalability. By treating verification as a layered, resource‑constrained process rather than a trivial mechanical step, the work reframes the trustworthiness of mathematical results as an emergent property of verifier ecosystems. The notion of unverifiability, introduced here, provides a formal vocabulary for discussing the inevitable limits of any verification regime, and the paper positions itself as a foundational citation for future work in theorem proving, software verification, and AI safety.