Generated Gates Inherit Their Generator's Blind Spots: A Diversity Floor for Self-Generated Verification
Authors/Creators
Description
A representation-theoretic account of why gates generated by a model family inherit that family's blind spots. Modeling a gate by the indistinguishability relation of its underlying family F — with stealth set Σ_F of faults the family cannot separate from acceptable artifacts — the paper proves: (GG1, inheritance) any gate constructed within F has a blind set containing Σ_F, so a generated gate cannot do better than the family it is built from; and (GG2, self-generation floor) a closed system that generates its own gates from a single founding family has an escape floor exactly λ_F, a fixed point invariant to generation depth, recursion, and prompt diversity. The floor is lowerable only by an exogenous distinction source: a genuinely different family, whose independent stealth set multiplies the floor down to λ_F·λ_N, or a reality-grounded gate that consults ground truth rather than any family's representation and is the unique non-representational escape. The results generalize the classical theory of coincident failures in N-version programming to the regime where the versions and their checkers are produced from within one representational family. Identifies the self-generation floor as the origin of the stealth mass that caps reliable horizon in long-horizon agentic work (companion: https://doi.org/10.5281/zenodo.20820968). Includes a self-contained Python simulation reproducing all numerical results in Table 1.
Files
paper_gg.pdf
Additional details
References
- D. E. Eckhardt and L. D. Lee. A theoretical basis for the analysis of multiversion software subject to coincident errors. IEEE Trans. Software Engineering SE-11(12):1511–1517, 1985.
- B. Littlewood and D. R. Miller. Conceptual modeling of coincident failures in multiversion software. IEEE Trans. Software Engineering 15(12):1596–1614, 1989.
- J. C. Knight and N. G. Leveson. An experimental evaluation of the assumption of independence in multiversion programming. IEEE Trans. Software Engineering SE-12(1):96–109, 1986.
- S. S. Brilliant, J. C. Knight, and N. G. Leveson. Analysis of faults in an N-version software experiment. IEEE Trans. Software Engineering 16(2):238–247, 1990.
- P. G. Bishop. The PODS diversity experiment. In Software Diversity in Computerized Control Systems, Springer, 1988.
- L. Hatton. N-version design versus one good version. IEEE Software 14(6):71–76, 1997.
- J. Ron, B. Baudry, and M. Monperrus. N-Version Programming with Coding Agents. arXiv:2606.20158, 2026.
- A. Panickssery, S. R. Bowman, and S. Feng. LLM Evaluators Recognize and Favor Their Own Generations. In Advances in Neural Information Processing Systems (NeurIPS), 2024. arXiv:2404.13076.
- E. Spiliopoulou et al. Family-level self-preference bias in language-model evaluators. 2025.
- Z. Ma, T. Zhang, M. Cao, J. Liu, W. Zhang, M. Luo, S. Zhang, and K. Chen. Rethinking Verification for LLM Code Generation: From Generation to Testing. In Advances in Neural Information Processing Systems (NeurIPS), 2025. arXiv:2507.06920.
- M. Thomas, T. Partridge, et al. Verification Limits Code LLM Training. arXiv:2509.20837, 2025.
- Pappu et al. Heterogeneous multi-agent teams fail to match their best member. arXiv:2602.01011, 2026.
- Z. Bugaud. Hidden Clones: Exposing and Fixing Family Bias in Vision-Language Model Ensembles. arXiv:2603.17111, 2026.
- R. Greenblatt, B. Shlegeris, K. Sachan, and F. Roger. AI Control: Improving safety despite intentional subversion. arXiv:2312.06942, 2024.
- M. A. Mohsin, M. Umer, A. Bilal, et al. On the Fundamental Limits of LLMs at Scale. arXiv:2511.12869, 2025.
- A. Scrivens. Information-Theoretic Limits of Safety Verification for Self-Improving Systems. arXiv:2603.28650, 2026.
- I. Matijašević. A Fault-Tolerance Threshold for Gated Agentic Computation: Reliable Long-Horizon Work from Unreliable Executors. 2026. Preprint, https://doi.org/10.5281/zenodo.20837102.