Work

Local Syntactic Self-Consistency and Global Semantic Defects: An Audit Framework for AI-Generated Formal Proofs in Arithmetic Geometry

Year: 2026
Type: preprint
Abstract: After combining large language models with formal proof assistants such as Lean 4, the output scale of lengthy mathematical candidate proofs is exceeding the processing capacity of traditional peer review. The formal kernel can rigorously verify whether symbol inference is closed in an axiomatic system, but it cannot automatically guarantee that the formal symbols faithfully correspond to the target mathematical object, nor can it automatically verify the globally consistent invariants required for infinite asymptotic arguments. Recently, AI formal research has repeatedly confirmed the same judgment: "compilation passed" does not mean "semantic fidelity", Lean code can compile successfully, but express an empty, incorrect, or weakened statement; The content of carrying capacity can be hidden in declarative axioms or local assumptions, and it is impossible to distinguish between a sound formalization and a formalization based on false or overly strong premises based solely on the statement 'it can be constructed without any sort'. The semantic fidelity problem in formal mathematics is therefore defined as an audit dimension independent of grammar validation.The goal of this article is to propose an independently verifiable audit framework for identifying semantic drift in AI generated arithmetic geometric proofs. The core distinction of the framework is the fidelity of deduction at the grammatical level and the isomorphism of interpretation from formal symbols to arithmetic geometric objects at the semantic level. The framework summarizes four types of high-frequency semantic drift patterns: proxy substitution, quantifier order reversal, Archimedean parsing component castration, and local global traps of bad prime union sets, and provides actionable audit anchors for each type. This article uses a simplified recursive error model to characterize the quantitative conditions for "finite slice autonomy and infinite limit phase transition collapse", and provides a minimum reproducible Lean 4+Mathlib audit example to enable quantifier consistency auditing to be implemented in the local environment. This article also refines the mathematical expressions involved in the framework itself, including the formal definition of the Riemann hypothesis, the fixed dual pair premise of the Rallis seesaw identity, the single group single truncation boundary of Arthur truncation, and the distinction between the representativeness of the Bost slope inequality and the accuracy of the original text. The criterion for testing a framework is whether it can expose corresponding semantic breaks on the smallest reproducible sample.
Author: Wu, Weilong
Language: English
Subfield: Software
Open Access status: green
OpenAlex ID: W7237346376
Cites:
Cited by:
Related to: