Congrats on the 100. Since you mentioned benchmark improvements, here is one from the board data itself.

ARC-Challenge is sitting below chance for most of the board, and it is doing real work in the ranking.

I pulled index.html and re-ran the numbers over all 176 models:

ARC-Challenge   mean 23.52   median 23.12   range 18.60 - 29.69
below chance (25):  139 of 176

A sign test on that split is z = 7.69 against a chance-level null. That one does not need any assumption about item count. Ability does not go systematically below chance, so a median 1.9 points under it usually means the option scoring is picking up a length or normalization bias rather than a skill.

The spread is the part that moves ranks. Between-model stdev on ARC-Challenge is 2.10. On the standard 1,172-item split a pure guesser already produces 1.27 of that from sampling alone, so the genuine between-model signal is about 1.7 points wide. ARC-Easy over the same 176 models is 7.33.

In getIntelligenceIndex the ARC term is (arc + arcChall) / 2, equal weight. So half of your widest-range benchmark is being averaged against a column with a quarter of the range that is centered below chance. At full coverage that is about 14% of the index.

What it costs: recomputing with ARC-Easy alone moves the average model 2.84 ranks, max 14 (Escarda-86M-Base 83 to 69), and reorders positions 3 through 5. Your top two hold either way.

I am not saying drop it. ARC-Challenge still has Spearman 0.555 with log params, so it is measuring something, and at 150M some models do clear chance (Zero-v0.1-150M at 28.84). It is the equal weighting I would question, not the benchmark.

Would you rather weight the composite by each column's usable range, or keep the weights readable and just show ARC-Easy and ARC-Challenge as separate columns?