analysis__draws10_t1_molmo2_40k_k4l2.json

{
 "arms_pooled": {
  "greedy": {
   "chunk_mae": 6.0079,
   "first_mae": 2.1871,
   "pred_key": "pred:bijou@40000"
  },
  "draws10_t1": {
   "chunk_mae": 5.8492,
   "first_mae": 1.9736,
   "pred_key": "pred:bijou@40000_draws10_t1"
  }
 },
 "subset_mode": false,
 "n_rows_paired": 25800,
 "read1_delta_ar": {
  "mean": -0.15422,
  "median": -0.06892,
  "ci95": [
   -0.19484,
   -0.11319
  ],
  "arm_win_rate": 0.5492,
  "n_frames": 17204,
  "most_influential_repos": [
   {
    "repo": "willnorris/bbox-2",
    "n_excl": 16,
    "mean_without": -0.14513
   },
   {
    "repo": "jlesein/TestBoulon7",
    "n_excl": 160,
    "mean_without": -0.14795
   },
   {
    "repo": "dopaul/1500_chess_moves",
    "n_excl": 600,
    "mean_without": -0.16002
   },
   {
    "repo": "Hale-Sage/merge_data",
    "n_excl": 260,
    "mean_without": -0.15958
   },
   {
    "repo": "yangfengzzz/yellow_block_grasp_vary",
    "n_excl": 40,
    "mean_without": -0.14978
   }
  ],
  "definition": "Delta_AR: paired per-frame chunk_mae, _draws10_t1 - greedy, core frames (negative = ensembling helps)"
 },
 "read2_fairness_vs_flow": {
  "delta_ar_mean": -0.15422,
  "flow_teacher_gain": -1.258,
  "ar_gain_smaller_than_flow": true,
  "definition": "Delta_AR vs the flow teacher's draws1 6.6232 -> draws10 5.365 gain (-1.258) on the same panel \u2014 does AR ensemble like flow, or is greedy already the mean?"
 },
 "read3_family_band": {
  "draws10_t1_chunk_mae": 5.8492,
  "flow_draws10_band": 5.365,
  "reaches_band": false,
  "anchors_record_only": {
   "teacher_draws10_heun30": [
    5.3645,
    1.4242
   ],
   "snapflow_student_mean10": [
    5.3675,
    1.5927
   ]
  },
  "subset_mode": false,
  "note": "full panel"
 },
 "read4_first_mirrors": {
  "delta_ar_first": {
   "mean": -0.21345,
   "median": -0.10407,
   "ci95": [
    -0.23451,
    -0.19243
   ],
   "arm_win_rate": 0.5887,
   "n_frames": 17204,
   "most_influential_repos": [
    {
     "repo": "HITHY/so100_peach4",
     "n_excl": 156,
     "mean_without": -0.20717
    },
    {
     "repo": "shin1107/koch_base_episodes",
     "n_excl": 64,
     "mean_without": -0.2072
    },
    {
     "repo": "shin1107/koch_new",
     "n_excl": 60,
     "mean_without": -0.20738
    },
    {
     "repo": "shin1107/koch_new2",
     "n_excl": 64,
     "mean_without": -0.20973
    },
    {
     "repo": "kevin510/lerobot-cat-toy-placement",
     "n_excl": 16,
     "mean_without": -0.2168
    }
   ],
   "definition": "first_mae mirror of read 1 (record-only)"
  },
  "flow_first_gain_descriptive": -0.5089,
  "flow_first_band_descriptive": 1.4242,
  "draws10_t1_first_mae": 1.9736
 },
 "read5_execution_oracles": {
  "row_pairing": "full byte-match",
  "state_copy_byte_match": true,
  "state_copy_norm_byte_match": true,
  "ar_temperature": 1.0,
  "sample_draws": 10,
  "draws_policy": "bijou@40000_draws10_t1",
  "checkpoint": "fontaine_molmo2_ar_40k_ddp4/step_040000",
  "report_crosscheck": "both arms reproduced (|d| < 5e-3)"
 },
 "expectations": {
  "e1_delta_negative": {
   "met": true,
   "ci95_excludes_zero": true
  },
  "e2_gain_smaller_than_flow": {
   "met": true
  },
  "e3_no_overtake_of_flow_band": {
   "met": true
  },
  "e4_falsifier": {
   "rule": "falsified if Delta_AR > +0.1",
   "falsified": false
  }
 },
 "decision": {
  "falsified": false,
  "assembly": [
   "E1 MET + E2 MET: Delta_AR < 0 with CI95 excluding 0 and |Delta_AR| < 1.258 \u2014 the AR gain is real but SMALLER than the flow teacher's; greedy decode already sits near the predictive mean (the mean-collapse shape)",
   "E3 as expected: _draws10_t1 5.8492 does not overtake the flow draws10 band 5.365"
  ]
 }
}