Interpolating Qwen2.5-0.5B and Qwen2.5-0.5B-Instruct with one global coefficient t leaves open whether per-layer coefficients help. We pre-registered that a coordinate search over four layer groups would beat the global t^* on validation, with a smaller advantage on test (the winner's curse). The search used 21 validation evaluations (1,199 items from four benchmarks) and 3,967 disjoint test items. The validation gain over t^*=0.7 is +0.83 pp, within noise (paired 95% CI [-0.58,+2.17]). On test the gain is +0.39 pp [-0.38, +1.15], p=0.311, not distinguishable from zero (no equivalence test); this is the primary comparison, as worded in hypothesis 5. Against the best original model (base, 0.5025) the layer-wise merge scores 0.5129, a gain of +1.04 pp [+0.20, +1.93] with nominal p=0.017. The pre-registered general rule (paired CI above zero) is met nominally, and fails only the Bonferroni threshold of 0.0083, added later, so we do not claim an improvement. The data are consistent with both a winner's curse and no shrinkage. Scope: one model pair, four benchmarks, one search run.
Download the PDF · Figshare record and reproducibility package