{"baseline":[{"num_evals":190,"policy":"dreaming_zebra","rank":1,"score":1735,"std":42.6},{"num_evals":745,"policy":"pi05_droid","rank":2,"score":1608,"std":30.7},{"num_evals":941,"policy":"pi0_fast_droid","rank":3,"score":1582,"std":29.6},{"num_evals":952,"policy":"paligemma_vq_droid","rank":4,"score":1546,"std":29.9},{"num_evals":949,"policy":"paligemma_diffusion_droid","rank":5,"score":1533,"std":29.7},{"num_evals":1057,"policy":"paligemma_fast_specialist_droid","rank":6,"score":1531,"std":29.8},{"num_evals":1062,"policy":"paligemma_fast_droid","rank":7,"score":1516,"std":29.9},{"num_evals":1120,"policy":"pi0_droid","rank":8,"score":1461,"std":29.8},{"num_evals":629,"policy":"paligemma_binning_droid","rank":9,"score":786,"std":50.0}],"cache_stale":false,"generated_at":"2026-10-11T20:45:24.866704Z","label":"nvidia","largest_change":{"base_evals":629,"base_rank":9,"base_score":786,"policy":"paligemma_binning_droid","rank_delta":1,"score_delta":-23,"without_evals":629,"without_rank":8,"without_score":763},"max_abs_rank_delta":1,"max_abs_score_delta":33,"official_policy_eval_threshold":100,"source":"official Bradley-Terry Davidson leaderboard rerun on counted A/B evals","top_changes":[{"base_evals":629,"base_rank":9,"base_score":786,"policy":"paligemma_binning_droid","rank_delta":1,"score_delta":-23,"without_evals":629,"without_rank":8,"without_score":763},{"base_evals":1062,"base_rank":7,"base_score":1516,"policy":"paligemma_fast_droid","rank_delta":1,"score_delta":-22,"without_evals":1019,"without_rank":6,"without_score":1494},{"base_evals":952,"base_rank":4,"base_score":1546,"policy":"paligemma_vq_droid","rank_delta":1,"score_delta":-21,"without_evals":913,"without_rank":3,"without_score":1525},{"base_evals":1120,"base_rank":8,"base_score":1461,"policy":"pi0_droid","rank_delta":1,"score_delta":-21,"without_evals":1076,"without_rank":7,"without_score":1440},{"base_evals":1057,"base_rank":6,"base_score":1531,"policy":"paligemma_fast_specialist_droid","rank_delta":1,"score_delta":-20,"without_evals":1019,"without_rank":5,"without_score":1511},{"base_evals":745,"base_rank":2,"base_score":1608,"policy":"pi05_droid","rank_delta":1,"score_delta":-18,"without_evals":700,"without_rank":1,"without_score":1590},{"base_evals":949,"base_rank":5,"base_score":1533,"policy":"paligemma_diffusion_droid","rank_delta":1,"score_delta":-14,"without_evals":909,"without_rank":4,"without_score":1519},{"base_evals":941,"base_rank":3,"base_score":1582,"policy":"pi0_fast_droid","rank_delta":1,"score_delta":-11,"without_evals":896,"without_rank":2,"without_score":1571}],"without_org":[{"num_evals":700,"policy":"pi05_droid","rank":1,"score":1590,"std":32.7},{"num_evals":896,"policy":"pi0_fast_droid","rank":2,"score":1571,"std":31.5},{"num_evals":913,"policy":"paligemma_vq_droid","rank":3,"score":1525,"std":31.9},{"num_evals":909,"policy":"paligemma_diffusion_droid","rank":4,"score":1519,"std":31.8},{"num_evals":1019,"policy":"paligemma_fast_specialist_droid","rank":5,"score":1511,"std":31.7},{"num_evals":1019,"policy":"paligemma_fast_droid","rank":6,"score":1494,"std":31.8},{"num_evals":1076,"policy":"pi0_droid","rank":7,"score":1440,"std":31.8},{"num_evals":629,"policy":"paligemma_binning_droid","rank":8,"score":763,"std":51.3}]}