This study evaluates LLM random number generation under a fully crossed 2×2×2 design: two models (claude-opus-4-6, gpt-5.4), two prompt styles (batch vs. sequential), and two number ranges (1–10, 0–255). Each of the eight conditionswas run 8 times with 125 numbers per run. Two findings are consistent across all conditions. First, batch mode produces a distributional-vs-sequential split: outputs pass KS uniformity tests while failing every independence test administered (ACFsignificant in all 8 runs for all batch conditions; no NIST SP 800-22 pass across both 0–255 batch conditions). Second, sequential mode produces prototype fixation: both models collapse to a single modal output (7 in the 1–10 range; 173 in the 0–255 range), with claude-opus-4-6 achieving complete fixation (std = 0) in the 1–10 condition. The cross-provider, cross-range convergence on specific prototype values distinguishes prototype fixation from generic distributional bias. No tested configuration produced output suitable for applications requiring both distributional uniformity and sequential independence.
Sutej Singh (Sun,) studied this question.