---
title: "N-version programming"
type: "entity"
entity_kind: "concept"
status: "hub"
canonical_name: "N-version programming"
aliases: ["NVP","N-version programming failure-independence assumption"]
first_seen: "2026-11-04T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["software fault tolerance","failure independence","generative-process diversity","correlated errors","algorithmic monoculture"]
seek_code_commit: "d40ee39"
---


N-version programming (NVP) is a software fault-tolerance technique, dating to the 1970s, that runs multiple independently-written implementations of the same specification in parallel and votes on their outputs — on the founding assumption that independently-produced implementations fail independently. That assumption has been empirically challenged since at least the Knight & Leveson studies of the 1980s, which found diversely-implemented programs still failing on overlapping inputs far more than chance would predict. The concept is commonly credited to Algirdas Avižienis, though this vault has not yet independently verified that attribution.

Matters to this vault as the decades-old ancestor theory that two 2026-07 preprints on LLM-generated code — [[claim-arxiv-2607-02808-03174-adjacent-not-corroborating-tieman-markou]] and [[claim-llm-generated-code-ensembles-realize-partial-nvp-reliability-gain]] — explicitly measure their own contribution against, and the same question in a different guise as the one [[claim-generative-process-diversity-predicts-lower-correlated-failure|Tieman & Markou (2026)]] ask about LLM output diversity and correlated failure: whether diversity among implementations actually buys the independence it's assumed to buy. Whether Tieman & Markou's own reference list engages the NVP literature at all is unchecked (see [[question-corroborate-tieman-markou-2026-generative-process-diversity]]).

- 2026-11-05: an independent (different author group, different institution, different statistic) replication of Knight & Leveson's original multi-version-independence test now exists for AI-agent-generated code — Ron, Baudry & Monperrus's "N-Version Programming with Coding Agents" (arXiv:2606.20158) reproduces the classic experiment with AI coding agents rather than direct LLM prompting, finding the same failure-independence violation arXiv:2607.02808 reports, via the Knight–Leveson z-statistic rather than a reliability-gain-fraction calculation. The original 1986 experiment this vault had cited without a full citation is Knight & Leveson, "An Experimental Evaluation of the Assumption of Independence in Multi-Version Programming," *IEEE Transactions on Software Engineering* 12(1):96–109, DOI 10.1109/TSE.1986.6312924. ([[claim-ron-baudry-monperrus-2026-independently-corroborate-llm-code-nvp-failure-independence-violation]])

## References
- [[claim-arxiv-2607-02808-03174-adjacent-not-corroborating-tieman-markou]]
- [[claim-nogueira-almeida-2026-llm-code-diversity-preprints-share-co-author]]
- [[claim-llm-generated-code-ensembles-realize-partial-nvp-reliability-gain]]
- [[claim-ron-baudry-monperrus-2026-independently-corroborate-llm-code-nvp-failure-independence-violation]]
- [[claim-nogueira-2026-0-43-0-44-figures-no-independent-corroboration-as-of-2026-11-05]]
- Capture: 10-inbox/raw/2026-11-04-a-dedicated-corroboration-search-for-tieman-markous-generative.md
- Capture: 10-inbox/raw/2026-11-05-does-any-independent-source-corroborate-nogueira-pattabiraman-vieira.md