This paper examines how benchmark results should be interpreted when applied to regulatory, viability-first cognitive architectures rather than optimization-driven systems. Using SpiralBrain v3.0 as an instrumented example, the study analyzes the Massive Multitask Language Understanding (MMLU) benchmark not as a measure of general capability, but as a structured stressor that probes internal system stability under uncertainty. The paper introduces cognitive integrity as a distinct evaluation axis, defined by boundedness, coherence, and recoverability of internal system dynamics. Through controlled MMLU stress tests, the results show that low benchmark scores (20–36%) can reflect intentional regulatory throttling to preserve homeostasis rather than deficiencies in reasoning capability. Internal integrity metrics—such as coherence, hazard, and drift—remain stable even as task accuracy is deliberately sacrificed. The work clarifies common misconceptions that equate benchmark accuracy with cognitive competence and argues for dual-axis evaluation: separating task-level performance from internal regulatory integrity. No claims of universality, superiority, or deployment readiness are made; conclusions are strictly limited to observed behavior in SpiralBrain v3.0. The study contributes to ongoing discussions in AI evaluation, safety, and interpretability by highlighting the limits of accuracy-only benchmarks for non-optimizing architectures.
John Cragin (Tue,) studied this question.