{
  "context": "BMVC paper Phase-2 review L24: bootstrap CI on Delta_contam = 9.03 pp (Run-C - Run-A on LVIS top-1).",
  "data_constraint": "Per-UID predictions are NOT available for Run-A or Run-C (only Run-0 baseline has them). Per-stack training logs underlie all numbers; only one of the five checkpoints on disk corresponds to a published 'Run-A' or 'Run-C' best LVIS number, namely Run-C T2-MLP (45.97). Run-A checkpoints were not preserved.",
  "headline_numbers_from_paper": {
    "delta_contam_single_stack_pp": 9.03,
    "delta_contam_two_three_stack_mean_pp": 9.67,
    "runC_two_stack_mean": 45.7,
    "runA_three_stack_mean": 36.03
  },
  "stack_level_bootstrap": {
    "method": "stack-level bootstrap of Delta_contam = Run-C - Run-A means",
    "runA_per_stack": [
      36.66,
      35.02,
      36.4
    ],
    "runC_per_stack": [
      45.43,
      45.97
    ],
    "n_runA_stacks": 3,
    "n_runC_stacks": 2,
    "point_estimate_pp": 9.673333333333332,
    "B": 10000,
    "bootstrap_mean_pp": 9.673100333333332,
    "SE_pp": 0.45816736958885046,
    "ci_95_percentile_pp": [
      8.85666666666667,
      10.68
    ],
    "ci_95_basic_pp": [
      8.666666666666664,
      10.489999999999995
    ],
    "p_delta_gt_0": 1.0,
    "p_delta_gt_5pp": 1.0,
    "welch_t_ci_95_pp": [
      7.778152531291753,
      11.568514135374912
    ],
    "welch_df": 2.8351417233594884
  },
  "eval_noise_bootstrap": {
    "method": "eval-noise bootstrap (within-run plateau, T2-MLP only, Run-A approximated)",
    "caveat": "Run-A metrics.jsonl not available; Run-A trajectory synthesized from Run-B-v2 trajectory (most similar prune-rate run) shifted to paper-reported best=35.02. Bounds eval-noise only; does NOT capture seed/init/mask-seed/data-order variance.",
    "runC_plateau_n": 85,
    "runC_plateau_mean": 45.43811784071081,
    "runC_plateau_sd": 0.20003729404781864,
    "runC_best": 45.9733784198761,
    "runA_synthetic_plateau_n": 60,
    "runA_synthetic_plateau_mean": 34.513706723054256,
    "runA_synthetic_plateau_sd": 0.28150240976628066,
    "runA_assumed_best": 35.02,
    "B": 10000,
    "best_minus_best": {
      "mean_pp": 10.929546449184413,
      "SE_pp": 0.05318238437225561,
      "ci_95_percentile_pp": [
        10.756429772377011,
        10.990169386863705
      ]
    },
    "mean_minus_mean": {
      "mean_pp": 10.92426089452265,
      "SE_pp": 0.042163262576233554,
      "ci_95_percentile_pp": [
        10.842112967699178,
        11.006345839395232
      ]
    }
  },
  "recommended_paper_report": {
    "headline": "Delta_contam = 9.67 pp [stack-level 95% CI given below]",
    "stack_bootstrap_ci_95_percentile_pp": [
      8.85666666666667,
      10.68
    ],
    "welch_t_ci_95_pp": [
      7.778152531291753,
      11.568514135374912
    ],
    "framing": "Use the stack-level CI as the headline. The eval-noise CI is reported as a sensitivity check.",
    "caveat_to_acknowledge": "Bootstrap is over the n=3 Run-A and n=2 Run-C training stacks, treating each stack as one independent realization of the 'final LVIS top-1' statistic. Sample sizes are small; the percentile bootstrap with replacement provides a conservative interval given the limited stack count."
  }
}