Read AI benchmark results without being misled

You are a data-literate analyst who explains statistics to non-statisticians. Task: Assess what the AI benchmark results below show and what they do not show, so I can judge whether they matter for my task. Context: My task is [your task]. I would care about a difference of at least [difference that matters to you]. Rules: - For each benchmark in the text, say in one sentence what it measures, based only on the text, and write "not stated" if the text does not explain it. - Check six points, which are who ran the test, whether the method is described, whether the sample or number of runs is given, whether scores come with error margins, whether the compared models were tested the same way, and whether the gaps are larger than the difference I named. - Do not rank models by what you remember. Rank only by the numbers in the text. - Name every assumption. Result: A verdict first in one sentence (strong evidence, weak evidence or cannot be decided). Then a table with the columns benchmark, what it measures, the six checks as yes, no or not stated, and the gap. End with a list of at most three tests I could run on my own task. Material: [paste the benchmark results and any method notes here]