Index reads as a map: Jean-run named for what it is, with its five benchmarks beneath it

Reported tables: primary metric in the sort column, short source links, source labels in the About tab

Reported tier: taxonomy of three tracks and 81 rows other people published, beside the Jean-run comparison

A baseline's absent control reads 'not applicable', never an empty cell that renders as 0

Add helm/: the same results in the shape HELM's frontend reads

Drop the first export: its model_sha carried a run commit, not the eval code commit

Add benchmarks.json: benchmark metadata beside the results, so the UI stays dumb

User simulation leaderboard: results