Auditing the Transferability of a Published Quantization Comparison: The Verdict Is a Function of the Slice

Practitioners choose between released 4-bit conversions of a model from published comparison sentences such as "AWQ consistently outperforms ... GPTQ ... (7B-70B)", although the files they serve, the serving stack and the way quality is measured all differ from what the sentence was written about. Later comparisons disagree about its direction without reporting intervals, so whether such a verdict transfers to deployment artifacts is unknown. We test the transfer on released artifact pairs by crossing the measurement regime (perplexity against downstream accuracy) with the serving engine on one 12 GB accelerator, and report each verdict with its interval, margin and realized precision. On an in-range 7B released pair from a family outside the audited table, WikiText-2 perplexity places AWQ behind by 0.993% while LAMBADA accuracy, chosen after reading eight tasks, places it ahead by 3.136 points, both beyond their margins on both engines; the opposition persists when one dataset is scored both ways. In a five-cell verdict matrix the measurement regime changed verdict states and the serving engine none; where the engines disagreed elsewhere (one audited-table checkpoint; BoolQ at 1.5B and 7B; LAMBADA on a preliminary second family), the intervals split between inconclusive and zero-excluded, and on the primary family's two pairs most tasks fixed before measurement, the primary one included, did not resolve their margin. The direction of a published quantization comparison can thus depend on the slice it is read on, so each verdict needs its slice and realized precision stated beside it.