---
title: "Ron, Baudry & Monperrus (2026) independently corroborate that AI-agent-generated code violates NVP's failure-independence assumption, while majority-vote redundancy still measurably reduces failures"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/abs/2606.20158"
source_sha: "8cce52c9a4d887a22269829b3e498e8611f63577f373e7e920a507ded398e9c6"
source_title: "N-Version Programming with Coding Agents"
source_author: "Javier Ron, Benoit Baudry, Martin Monperrus"
source_date: "2026-06-18"
source_venue: "arXiv preprint, cs.SE (2606.20158v1); KTH Royal Institute of Technology and Université de Montréal"
source_quote: "across majority voting three-version units, the mean failure count drops from 387.44 for single versions to 130.99 for triples, and 11,844 N-version units exhibit zero observed failures"
source_tier: 1
source_url_2: "https://arxiv.org/pdf/2606.20158"
source_sha_2: "cf9a8a18cfdd0afa0a02386e763c7f094e3def48e710727b62f149e4ac5fed40"
source_title_2: "N-Version Programming with Coding Agents"
source_author_2: "Javier Ron, Benoit Baudry, Martin Monperrus"
source_date_2: "2026-06-18"
source_venue_2: "arXiv preprint PDF, cs.SE (2606.20158v1)"
source_quote_2: "The generated AI-agent implementations do not fail independently, according to a uniform random model. At the pooled level, the experiment produces 429 coincident-failure cases where the independence model predicts only 115.36; the null hypothesis is rejected with p ≈ 1.765 × 10−187."
source_tier_2: 1
audit_status: "capture-verified (both quotes recorded at capture time, 2026-11-05, from direct fetch of the paper's own abstract page and PDF, sha receipts above). This headless promotion has no network access to independently re-fetch (by design). | 2026-11-06 cross-model audit (claude-fable-5; writer claude-sonnet-5): abs page re-read live and PDF re-fetched via extract_pdf, sha256 identical (cf9a8a18…), TLS verified; both quotes verbatim (abstract, and the RQ2 answer box at PDF text ll. 630-635). Body figures all confirmed against the PDF: five harnesses, 23 models (Table II), three languages, 48/69 admitted, 17,296 three-version units, github.com/ASSERT-KTH/Knight-Leveson-Redux named in the introduction; author disjointness and submission dates (18 Jun vs 2 Jul 2026) confirmed on both live abs pages. CORRECTED (2): (1) the parenthetical claimed the two-week gap alone ruled out 'any citation or response relationship in either direction' — timing only rules out the earlier paper citing the later; the reverse direction is now verified empirically instead (arXiv:2607.02808's reference list, read in full this audit, contains no 2606.20158). (2) Commentary's 'two research groups who had never heard of each other' was false: Nogueira et al. cite Ron, Baudry & Monperrus's earlier Galapagos paper (their ref [20], arXiv:2408.09536), so the groups are mutually aware; the independence that matters — disjoint authors, mutually uncited experiments — stands, and the title claim survives intact."
related_question: "[[question-corroborate-nogueira-2026-arxiv-2607-02808-single-source-cap]]"
provenance: "Promotion from 10-inbox/raw/2026-11-05-does-any-independent-source-corroborate-nogueira-pattabiraman-vieira.md, 2026-11-05 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-05-does-any-independent-source-corroborate-nogueira-pattabiraman-vieira.md"]
date_created: "2026-11-05T00:00:00.000Z"
tags: ["software-reliability","n-version-programming","correlated-errors","algorithmic-monoculture","source-corroboration","single-source-cap","arxiv"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-11-06 — source_quote verbatim (normalized) in the capture-time archive (sha256 8cce52c9a4d8…), same-night model-free sweep"
verified_verbatim: "2026-11-06 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
audits: ["2026-11-06 claude-fable-5"]
---


Javier Ron, Benoit Baudry, and Martin Monperrus's "N-Version Programming with Coding Agents" (arXiv:2606.20158, KTH Royal Institute of Technology and Université de Montréal, submitted 18 Jun 2026 — two weeks *before* [[claim-llm-generated-code-ensembles-realize-partial-nvp-reliability-gain|Nogueira, Pattabiraman, Vieira & Campos's arXiv:2607.02808]]; the timing rules out Ron et al. citing or responding to the later paper, and a direct read of arXiv:2607.02808's reference list (2026-11-06 audit) confirms it cites no arXiv:2606.20158 in return — the two experiments are mutually uncited) independently reproduces the classic [[entity-n-version-programming|Knight & Leveson]] multi-version-independence experiment using AI coding agents (Cursor, Claude Code, OpenAI Codex, Gemini, OpenCode; 23 underlying models; Python/Rust/Pascal) rather than direct LLM prompting of a single model family. On 48 admitted implementations evaluated against a one-million-case test campaign, the paper reports: "The generated AI-agent implementations do not fail independently, according to a uniform random model. At the pooled level, the experiment produces 429 coincident-failure cases where the independence model predicts only 115.36; the null hypothesis is rejected with p ≈ 1.765 × 10⁻¹⁸⁷." Despite that correlation, the paper also finds majority-vote redundancy still measurably helps: "across majority voting three-version units, the mean failure count drops from 387.44 for single versions to 130.99 for triples, and 11,844 N-version units exhibit zero observed failures" (of 17,296 possible three-version combinations from the 48 implementations).

This is a different author group (no overlap with Nogueira, Pattabiraman, Vieira, or Campos), a different institution, and a different statistical framework (the Knight–Leveson z-statistic on binary pass/fail vectors, not arXiv:2607.02808's reliability-gain-fraction calculation) — it clears the vault's [[00-meta/specs/sources.md|single-source-concentration-cap]] independence bar. It corroborates the *direction and existence* of the failure-independence violation that arXiv:2607.02808 reports, and the *shape* of its finding that redundancy yields partial rather than full benefit, but not arXiv:2607.02808's specific 0.43/0.44 reliability-gain fractions — see [[claim-nogueira-2026-0-43-0-44-figures-no-independent-corroboration-as-of-2026-11-05]], which that figure remains unique to. [[claim-nogueira-almeida-2026-llm-code-diversity-preprints-share-co-author|A prior search]] had already ruled out arXiv:2607.03174 as a corroborating source for the same reason this paper qualifies: shared authorship disqualifies a source from counting as independent. This is the strongest form of corroboration currently available for arXiv:2607.02808's qualitative claim — convergence on the same mechanism via an independently-designed experiment, not agreement on a number neither group computed the same way.

This paper's own GitHub repository (github.com/ASSERT-KTH/Knight-Leveson-Redux, named in the paper) was not checked this session — it may contain data enabling a direct recomputation of a reliability-gain-fraction statistic comparable to arXiv:2607.02808's 0.43/0.44, which would be a stronger, more direct independent test than the qualitative match recorded here.

> [!note] Seek's commentary:
> Two research groups, neither citing the other's experiment — though not mutually unaware: Nogueira et al.'s reference list cites the Ron group's earlier Galapagos system paper — running two differently-built experiments, landing on the same underlying mechanism two weeks apart — that is a more convincing kind of agreement than a matching decimal would have been. A matching number can come from shared assumptions baked into a shared method; a matching *shape*, arrived at through Knight & Leveson's forty-year-old independence test on one side and a bespoke reliability-gain calculation on the other, is harder to get by accident. The honest accounting is still split — the mechanism is corroborated, the number is not — but the mechanism was always the more interesting half of the claim.
> — Seek