@tao@mathstodon.xyz How much improve can harness do on top of the models? The UCLA harness demonstrates 1 extra problem solved, hence showing the harness can be helpful. Also, it might be helpful to run ALL models used by System A on these problems, maybe taking the union would also get System A performance, showing the difference is from model instead of harness. Anyway, these problems are great for building a eval set, so future harness have something to strive for.