{
 "npz": "reports/eval__molmoact2_so101_released__panel_v2_k4l2_euler10_draws1_stable.npz",
 "checkpoint": "/home/ubuntu/checkpoints/molmoact2-so101-released",
 "wear_facts": {
  "leg_molmo_norm": "checkpoint",
  "worn_table": {
   "mean": [
    3.343996486826433,
    -35.7905980370996,
    30.20220128113388,
    55.88144220174933,
    -11.543010633027725,
    11.25886240824774
   ],
   "std": [
    28.909870406169997,
    52.25069634659296,
    47.94432906599221,
    36.01019142727721,
    69.35504013212369,
    17.116239869449775
   ],
   "q01": [
    -42.1300246338976,
    44.81741641835005,
    -54.59940817037187,
    4.929781836327758,
    -65.57568617645342,
    -0.3016556932619033
   ],
   "q99": [
    48.55349563198916,
    -96.10646680077767,
    83.60767220139971,
    93.41056417929472,
    43.53107398260694,
    44.74649336930881
   ]
  },
  "verdict": "the leg wore the released checkpoint's own q01q99 global table (molmo_norm=checkpoint); this re-expression moves the OUTPUT side onto the same honest per-repo rows the disc-1000 27.40 reference wears"
 },
 "caveats": [
  "output-side only: the state input stayed normalized/binned through the released table when the model ran \u2014 this bounds the wear share, it does not re-run the model",
  "per-repo rows are fit on the panel's own truth (oracle-ish rows, optimistic vs a deployment estimate) \u2014 identical estimator and panel as the disc-1000 audit, which is the point: same wear on both sides"
 ],
 "anchors_reproduced": {
  "pred:bijou@molmoact2-so101-released": 25.892388234103436,
  "state-copy": 8.367757873526733
 },
 "round_trip_worst_abs_deg": 1.52587890625e-05,
 "honest_wear_chunk_mae": 27.135121955707675,
 "honest_wear_per_joint": {
  "shoulder_pan": 12.272583961486816,
  "shoulder_lift": 66.08007049560547,
  "elbow_flex": 36.19916915893555,
  "wrist_flex": 17.373897552490234,
  "wrist_roll": 23.928756713867188,
  "gripper": 6.94985294342041
 },
 "own_table_per_joint": {
  "shoulder_pan": 6.443148612976074,
  "shoulder_lift": 68.90454864501953,
  "elbow_flex": 43.08244705200195,
  "wrist_flex": 12.39787769317627,
  "wrist_roll": 17.627012252807617,
  "gripper": 6.887579917907715
 },
 "null_repo_midpoint_chunk_mae": 25.154475696265422,
 "n_repo_rows": 838,
 "per_repo_worst10": [
  {
   "repo": "sihyun77/mond_13",
   "elements": 480,
   "own_table_mae": 32.26658248901367,
   "honest_wear_mae": 47.20325469970703
  },
  {
   "repo": "liuhuanjim013/so100_th_1",
   "elements": 2432,
   "own_table_mae": 32.68760681152344,
   "honest_wear_mae": 44.549095153808594
  },
  {
   "repo": "duthvik/sputnik_100_31_pour_liquid",
   "elements": 236,
   "own_table_mae": 50.8201904296875,
   "honest_wear_mae": 44.50764083862305
  },
  {
   "repo": "sihyun77/suho_3_17_1",
   "elements": 360,
   "own_table_mae": 22.740739822387695,
   "honest_wear_mae": 43.544532775878906
  },
  {
   "repo": "liuhuanjim013/so100_block",
   "elements": 1736,
   "own_table_mae": 36.63359069824219,
   "honest_wear_mae": 43.285682678222656
  },
  {
   "repo": "pengjunkun/so100_push_to_hole",
   "elements": 568,
   "own_table_mae": 23.262035369873047,
   "honest_wear_mae": 42.64848327636719
  },
  {
   "repo": "sihyun77/suho_3_17_3",
   "elements": 570,
   "own_table_mae": 28.55667495727539,
   "honest_wear_mae": 42.58462142944336
  },
  {
   "repo": "liuhuanjim013/so100_th",
   "elements": 840,
   "own_table_mae": 32.63666915893555,
   "honest_wear_mae": 42.24125289916992
  },
  {
   "repo": "masato-ka/so100_nlact_block_instruct_v3",
   "elements": 1043,
   "own_table_mae": 35.264835357666016,
   "honest_wear_mae": 41.521385192871094
  },
  {
   "repo": "kantine/domotic_vegetagblesAndFruitsSorting_anomaly",
   "elements": 240,
   "own_table_mae": 23.771739959716797,
   "honest_wear_mae": 41.241722106933594
  }
 ],
 "same_wear_comparison": {
  "released_own_table": 25.892392784479092,
  "released_honest_wear": 27.135121955707675,
  "sft_disc1000_honest_wear": 27.395334001919885,
  "delta_sft_minus_released": 0.26021204621221017,
  "null_repo_midpoint": 25.154475696265422,
  "read": "wear held fixed (honest per-repo rows on both sides), the SFT checkpoint ends within noise of where it started, and both sit AT/below the repo-midpoint null \u2014 SFT neither destroyed nor built community competence"
 }
}