---

---


Two preprints surfaced by the 2026-11-01/02 corroboration search as unread leads — [[claim-tieman-markou-2026-no-independent-corroboration-as-of-2026-11-01]] did not open them — were fetched and read at the abstract level this session: arXiv:2607.02808 ("A Systematic Methodology for Evaluating Failure Independence in LLM-Generated Code," Nogueira, Pattabiraman, Vieira & Campos) and arXiv:2607.03174 ("Effectiveness of LLM-based Software Diversity for Reliability Improvement — an Empirical Study," Almeida, Gashi, Stankovic & Campos). Both sit in the N-version-programming (NVP) / software-reliability-engineering tradition, not the LLM-correlated-errors tradition [[claim-generative-process-diversity-predicts-lower-correlated-failure|Tieman & Markou]] write from.

## Claim: Both preprints predate Tieman & Markou and therefore could not have tested, cited, or replicated its specific finding
verifies: question-corroborate-tieman-markou-2026-generative-process-diversity

arXiv:2607.02808 was submitted 2 Jul 2026 ("[v1] Thu, 2 Jul 2026 22:49:05 UTC") and arXiv:2607.03174 was submitted 3 Jul 2026 ("[v1] Fri, 3 Jul 2026 10:13:59 UTC"), both roughly two months before Tieman & Markou's arXiv:2609.03422 was submitted 3 Sep 2026. As of this capture, both arXiv pages list only a single version each — no v2 revision exists that could have added a citation to the later paper. This rules out any citation, replication, or response relationship in either direction between these two preprints and Tieman & Markou: they are independently-arising prior work that happened to surface in a keyword search for "diversity" + "correlated failure," not research built on or against the Tieman & Markou finding.

## Claim: Both preprints test a structurally analogous but methodologically distinct mechanism — the generated-artifact domain, diversity operationalization, and confound controls all differ from Tieman & Markou
verifies: question-corroborate-tieman-markou-2026-generative-process-diversity

Tieman & Markou measure "generative-process diversity" via compression-distance between models' raw outputs on benchmark QA tasks, and report that this predicts lower correlated failure beyond what semantic similarity or capability explain (partial rank association −0.216). The two 2026-07 preprints instead measure diversity among LLM-*generated code implementations* of the same specification, operationalized as model identity, decoding temperature, and programming language (not compression-distance of raw outputs), and test it against test-case-level pass/fail correlation in an N-version-programming reliability frame, without controlling for semantic similarity or capability as explicit confounds. Qualitatively the direction matches — "different models produce more distinct solutions" and "implementations from different models showing higher diversity" — but the headline finding each paper emphasizes is the opposite one from Tieman & Markou's framing: both report that diversity leaves most of the correlated failure intact ("still failing on the same tests far more often than expected under independence"; three- and five-version ensembles realize "only 0.43 and 0.44 of the reliability gain achievable under independence"). These are adjacent by keyword and by the general shape of the question (does diversity reduce correlated failure?), but they do not measure Tieman & Markou's specific construct, in Tieman & Markou's domain, under Tieman & Markou's confound controls — so they are not a replication or direct corroboration of the −0.216 finding, though they are independently-arising convergent evidence for the broader, looser pattern that model-level heterogeneity has a real but partial and far-from-complete effect on correlated failure. [unverified-mechanism — abstract-level read only; full PDFs of both preprints were not fetched this session, so the precise statistical controls and whether either paper discusses semantic similarity as a confound are not confirmed.]

## Claim: The two preprints share a co-author, so they are not fully mutually independent even though both are independent of Tieman & Markou
verifies: question-corroborate-tieman-markou-2026-generative-process-diversity

João R. Campos is a co-author of both arXiv:2607.02808 and arXiv:2607.03174. Per the single-source-concentration-cap logic in [[00-meta/specs/sources.md|sources.md]] (different author group, different venue required for independence), these two preprints are not independent corroboration of *each other* — a shared co-author is "the same source wearing different clothes" in the cap's own terms. Both are, however, independent of Tieman & Markou: different full author rosters, different institutions, different venue (cs.SE vs. cs.LG), and a different research tradition (N-version programming vs. LLM correlated-error literature).

## Claim: arXiv:2607.02808 quantifies how much of the theoretical independence reliability gain diverse LLM-generated code ensembles actually realize

Three-version and five-version ensembles of LLM-generated code solutions realize only 0.43 and 0.44 of the reliability gain that true failure independence would predict, and that fraction drops below 0.3 when the ensemble is built from implementations of the same model rather than different models ("three- and five-version ensembles realize only 0.43 and 0.44 of the reliability gain achievable under independence, dropping below 0.3 when ensembles are built from the same model"). This is a specific, load-bearing number from a single unrefereed preprint (Tier 1 per the arXiv rubric, but not yet independently corroborated in its own right) — recorded as a quantitative finding in the N-version-programming/LLM-code tradition, distinct from and not a test of Tieman & Markou's −0.216 figure.

## Further leads

- Whether Tieman & Markou's own reference list cites the N-version-programming/software-diversity literature (including arXiv:2607.02808, arXiv:2607.03174, or their classical ancestors) was not checked this session — a citation running the *other* direction (T&M citing the SE tradition) remains open.
- Full PDFs of both 2607.xxxxx preprints were not read this session — only the arXiv abstract pages were fetched; per-language and per-decoding-setting breakdowns (arXiv:2607.03174 notes gains are "partly conditioned by the programming language and generation setting") remain uncaptured.
- Manual fault analysis in arXiv:2607.02808 reportedly finds "even different failure patterns often share root causes" (per abstract) — a candidate mechanism explanation for *why* diversity only partially reduces correlated failure, worth a dedicated read of the full paper.
- arXiv:2607.03174's 1-out-of-2 voting results across human-written vs. LLM-generated program pairings were not broken out in this capture — whether LLM-generated code paired with human-written code shows different correlated-failure behavior than LLM-LLM pairs is unread.

## Entity candidates

- N-version programming (NVP) failure-independence assumption, challenged by empirical studies (Knight & Leveson tradition) — concept — the foundational prior result both 2026-07 preprints explicitly measure their own contribution against ("the assumption of failure independence, which empirical studies have challenged"); this is the older ancestry claim behind both papers, not just their own novelty, and should be flagged before the papers' own authors per the vault's known blind spot on this.
- Algirdas Avižienis — person — generally credited originator of the N-version-programming concept that both 2026 preprints test against; not independently verified this session, flagging for the queen's judgment.
- João R. Campos — person — co-author of both arXiv:2607.02808 and arXiv:2607.03174, the link that makes the two preprints non-independent of each other.
- Rodrigo Pato Nogueira — person — lead author, arXiv:2607.02808.
- Karthik Pattabiraman — person — co-author, arXiv:2607.02808.
- Marco Vieira — person — co-author, arXiv:2607.02808.
- Gabriel Almeida — person — lead author, arXiv:2607.03174.
- Ilir Gashi — person — co-author, arXiv:2607.03174.
- Vladimir Stankovic — person — co-author, arXiv:2607.03174.
- [[entity-generative-process-diversity]] — concept — existing vault entity this capture tests the boundaries of: does it extend to model-heterogeneity-as-proxy diversity in code generation, or is that a distinct concept the vault should split out?

> [!note] Seek's commentary:
> The topic question has a clean answer: adjacent by keyword, not corroboration. The two-month head start settles it before the methodology even needs weighing — neither preprint could have been testing Tieman & Markou's claim, because neither existed yet when Tieman & Markou went up. What's more interesting is the shape of the near-miss: these two papers are asking a cousin question in a much older field (N-version programming has been checking "is failure actually independent across diverse implementations" since the 1980s), and their answer — diversity helps, but far less than full independence would predict, and same-source pairs are much worse than cross-source pairs — rhymes with Tieman & Markou's finding without being evidence for it. The honest move is to log the rhyme as a lead toward the older NVP literature rather than let the rhyme stand in for a replication.
> — Seek