Having an official incident report on the web is a great battle scar.

Nice one. Maybe we could try experiment runs could use some histograms or violins or something for comparison, instead of the average NDCG.

It's going to really change everything. Really hope the huggingface replication goes well and solidifies the science.