Parent-Conflict Arbitration for Near-Tied Model Merges

A model merge carries a comparison instrument that ordinary checkpoint selection lacks: its parent checkpoints reveal where the specialized behaviors being composed conflict. The Frozen Parent-Conflict Audit (FPCA) turns this composition history into selective arbitration. Before candidate scoring, it measures conflict coverage and fixes answer-preserving neighborhoods, references, generators, and thresholds; it then compares neighborhood reference risk (NRR) only when the Global gap is at most 0.2 points and its paired interval contains zero. In a predeclared screen of 160 comparisons from Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct task-vector collections, 46 pairs enter this regime and FPCA selects a different candidate from the Global point estimate in 30, including 16 of 27 construction-only near-ties. A separately authored evaluator spanning 24 stratified cases and 9,600 exact-key prompts confirms the decision consequence: FPCA lowers mean NRR from 15.2 to 12.9 (paired reduction 2.3; 95% bootstrap CI [1.4,3.2]) with 16/3/5 wins/ties/losses. Equal-budget slices isolate the parent-conflict signal in two detailed cases, while conflict sensitivity identifies risk-controlling blocks in all six model–regime settings and enables a 0.5%-residual refinement that improves NRR by 1.3–1.7 points over the strongest matched-budget non-conflict targeting controls. Parent behavior therefore supplies actionable selection information precisely where aggregate accuracy leaves a merge decision unresolved.