Weight quantization reduces parameter storage, but its effect on answer selection need not be summarized by candidate-level classification accuracy. We report a controlled local pilot using Qwen2.5-0.5B-Instruct as a prompted arithmetic judge at native BF16, 8-bit and 4-bit weight precision. The experiment evaluates 120 constructed questions, with 40 used for threshold development and 80 held out; each contains 16 distinct integer answers with exactly one correct answer. At N=16, selected-answer accuracy is 13.8% at native precision, 12.5% at 8 bits, and 7.5% at 4 bits. The 4-bit versus native paired difference is -6.2 percentage points, with a 95% question-bootstrap interval of [-15.0, 1.2]. We additionally evaluate development-calibrated selective rescoring. Native 0.5B batching changes N=16 selections on 28.7% of test questions, revealing substantial implementation sensitivity. A subsequent 1.5B same-family extension on the same questions yields N=16 accuracies of 28.7% (native), 28.7% (8-bit), and 20.0% (4-bit). This study reports actual quantized-model inference over artificial answer pools, not generated reasoning traces. It establishes neither a general scaling law nor a novel state-of-the-art method.
