Jas: AI-Paired Engineering as a Revival of N-Version Programming

A single developer built five working ports of a vector illustration application across Rust, Swift, OCaml, Python, and browser platforms in roughly 120 evening hours using AI-paired engineering, according to a case study posted to arXiv [1]. The five ports share a 23,000-line executable YAML specification that serves as the single source of truth, while per-port native code ranges from 0 to approximately 95,000 lines, reflecting what the author calls the specification's escape hatch [1]. The methodology pairs AI-assisted implementation with two safeguards: the precise specification and parallel implementations that function as a built-in differential-testing layer [1]. The author argues that AI-paired engineering, conditional on these two safeguards, makes feasible a scope of work that conventionally requires multiple developer-years [1]. The paper frames the approach as a revival of N-version programming, a technique explored in the 1980s but abandoned on cost grounds that AI now changes [1]. N-version programming originally proposed developing multiple independent implementations of the same specification to detect faults through comparison, but the expense of maintaining parallel codebases proved prohibitive for most commercial software projects. The case study reports concrete artifacts alongside what the author describes as honest limitations of the single-developer approach [1]. The paper does not claim generalizability across larger teams or different software domains, and the author notes that the results reflect one developer's experience working in evening sessions rather than a controlled experiment [1]. The work was submitted to arXiv on June 5, 2026, under the Computer Science category for Software Engineering [1]. The paper is titled "Jas: AI-Paired Engineering as a Revival of N-Version Programming" and is available as both PDF and HTML on the preprint server [1]. The author frames the contribution as a case study rather than a formal validation, emphasizing that the methodology's effectiveness depends on maintaining both the executable specification and the parallel-implementation testing layer [1].