{
 "values": {
  "gs1_executed_group_size": 1.0,
  "rollouts_per_step_both_arms": 64.0,
  "gs1_effective_contributors_median": 57.710074358176044,
  "gs1_mean_abs_baseline_minus_uniform_loo": 0.02577038449071199,
  "batch_mean_reward_pilot": 0.6130208333333335,
  "executed_base_model": "Qwen/Qwen3.5-2B"
 },
 "checks": {
  "gs1_executed_group_size": true,
  "rollouts_per_step_both_arms": true,
  "gs1_effective_contributors_median": false,
  "gs1_mean_abs_baseline_minus_uniform_loo": true,
  "batch_mean_reward_pilot": true,
  "executed_base_model": true
 },
 "passed": false,
 "ranges": {
  "gs1_executed_group_size": [
   1,
   1
  ],
  "rollouts_per_step_both_arms": [
   64,
   64
  ],
  "gs1_effective_contributors_median": [
   2,
   40
  ],
  "gs1_mean_abs_baseline_minus_uniform_loo": [
   0.02,
   null
  ],
  "batch_mean_reward_pilot": [
   0.1,
   0.95
  ]
 },
 "extra": {
  "grpo_batch_mean_reward_pilot": 0.5953125,
  "gs1_n_eff_per_step_median": [
   57.710074358176044,
   58.28483985404487,
   57.5644523852564
  ],
  "gs1_frac_correct_per_step": [
   0.59375,
   0.59375,
   0.71875
  ],
  "grpo_frac_correct_per_step": [
   0.734375,
   0.6875,
   0.4375
  ]
 },
 "b_eff_diagnostic": {
  "n_tasks": 24,
  "rollouts_per_task": 8,
  "sigma2_within": 0.16740327380952383,
  "replays": 2000,
  "seed": 0,
  "uniform_loo": {
   "mse_corrected": 0.11797014436492934,
   "B_eff": 14.078671367255739
  },
  "gs1_tfidf": {
   "mse_corrected": 0.11727581629968373,
   "B_eff": 14.113009037254281
  },
  "note": "batch B=T calibration tasks, one rollout each; target = mean of other 7 same-task rollouts; noise sigma2/7 subtracted"
 }
}