This preprint reports 180 controlled evaluations across six frontier AI models (Claude Sonnet 4.5, ChatGPT GPT-5.2, Grok 4/4.1, DeepSeek V2.5, Gemini 3, Perplexity), three independent evaluation runs, and ten cross-domain reasoning problems spanning eight scientific disciplines (applied physics, limnology, clinical biostatistics, cardiovascular physiology, evolutionary biology, atmospheric chemistry, Bayesian statistics, evolutionary anthropology, fluid dynamics, and game theory/ecology). The central finding is a 14-percentage-point reliability downturn relative to Study 1 formal domains (79.2% vs 93.0%), placing semi-determinate performance in Regime 2 (mixed stochastic-systematic error structure). Performance stratifies into a reliable core (Q3, Q6, Q7, Q8: 100% across all 18 evaluations) and a variable periphery determined by depth of mechanistic recall required. Critically, self-audit calibration (Spearman ρₛ) emerges as an independent architectural design variable: only Claude (ρₛ = 0.903) and Gemini (ρₛ = 0.703) meet the proposed Auditor-quality threshold. Kruskal-Wallis confirms significant model-level differentiation (H = 12.779, p = 0.026). This is Empirical Study 2 of a planned three-study programme. Study 1 covered formal determinate domains (mathematical reasoning). Study 3 will extend to fully indeterminate domains. Files included:- Main manuscript (PDF)- Supplementary Data S2: Complete 180-evaluation dataset, all model responses across 3 runs, self-audit scores, logic trail analysis, inter-rater reliability data, and scoring rubrics (PDF)
Pandit et al. (Sun,) studied this question.