Large language models (LLMs) perform strongly in machine translation, but their effectiveness as translation refiners depends on the draft source. We compare direct translation, self-refinement, and cross-system refinement on English-to-Chinese WMT test sets. Sockeye drafts obtain higher lexical-overlap scores but lower COMET than the stronger Qwen3 drafts, and receive larger gains after Qwen3 refinement. Additional controls qualify this pattern. Qwen3-0.6B drafts obtain the largest raw gain but lower final quality, while Qwen3-1.7B drafts start near Sockeye in COMET yet benefit less. Cross-family GLM-4-9B-Chat drafts also improve with Qwen3-8B, but achieve smaller gains and lower final quality than refined Sockeye drafts. Descriptive bucket, common-support, and regression analyses suggest that initial quality alone is an incomplete account of these differences. xCOMET-XL score and span diagnostics further characterize the contrast with Qwen3-8B self-refinement. We frame translation editability as a property of the evaluated draft–editor pairing, assessed jointly through quality gain, improvement probability, degradation risk, and final quality.