Each cell is the mean divergence over repeated independent samples, scored by an LLM judge against a fixed English reference answer for the same question. Bars left of the red line are inside the noise floor: the same drift you get from running the English prompt twice, which is non-zero because inference is not deterministic even at low temperature.
The judge shares a model family with the systems it scores, so it inherits their priors and will under-report drift on exactly the dimensions where those priors are doing the work. Treat the ordering as more reliable than the absolute values. The raw generation, the back-translation and the judge's one-line delta for every cell are carried in this page - expand any language row to read them. The dataset, the generator that produced it and the source of this page are in the public repository.