{
  "accessibility": {
    "article_alt_text_check": "scripts/check_site.py compares the article shortcode alt text with this receipt",
    "color_is_not_the_only_channel": true,
    "direct_title_and_description": true,
    "full_text_equivalent": "receipts/per-query.csv",
    "svg_role_img": true
  },
  "alt_text": "Sorted query-level SciFact nDCG at 10 differences for late minus naive chunking: 49 queries improved, 229 tied, and 22 worsened. The mean difference is +0.0196, with a paired query-bootstrap interval from +0.0033 to +0.0358. Points above zero favor late chunking; points below zero favor naive chunking.",
  "claim_scope_exclusions": [
    "A positive mean does not imply every query improves.",
    "The paired differences do not establish a universal effect."
  ],
  "data_source": {
    "aggregation": "query-level metrics aggregated by arithmetic mean",
    "artifacts": [
      {
        "path": "receipts/aggregate.json",
        "sha256": "71bafc1d81fc24ac1d30104e5f28af9b43142430b3ce798b10bd1061f1abb181"
      },
      {
        "path": "receipts/per-query.csv",
        "sha256": "377e6d4feb59687d5f3e0557dea05f4cae6de4dc08619d6069cab9ef782c487d"
      }
    ],
    "selection": "all 300 queries in the fixed SciFact test qrels",
    "transformation": "Subtract naive nDCG@10 from late nDCG@10 within each query, then sort the 300 paired differences for plotting; compute the displayed mean and paired percentile-bootstrap interval from those differences."
  },
  "description": "Sorted paired query-level nDCG@10 differences for late minus naive chunking.",
  "figure_id": "late-chunking-scifact-query-deltas",
  "plotted_data": {
    "improved_queries": 49,
    "markers": {
      "improved": "circle",
      "tied": "diamond",
      "worse": "square"
    },
    "mean_difference": 0.019552175132635588,
    "query_bootstrap_95_percent_interval": [
      0.0032721920785872076,
      0.03584544477819499
    ],
    "sorted_late_minus_naive_ndcg_at_10": [
      -0.698970004336,
      -0.666666666667,
      -0.643792812892,
      -0.569323441927,
      -0.369070246429,
      -0.369070246429,
      -0.369070246429,
      -0.369070246429,
      -0.289064826318,
      -0.210935173682,
      -0.143792812892,
      -0.130929753571,
      -0.130929753571,
      -0.113147192765,
      -0.080279210852,
      -0.055177191444,
      -0.044268507015,
      -0.030645620127,
      -0.022873853775,
      -0.014434881122,
      -0.014434881122,
      -0.014381337571,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.004058109292,
      0.022873853775,
      0.025257047143,
      0.029718496618,
      0.030645620127,
      0.034946891534,
      0.043823750839,
      0.045660685754,
      0.055177191444,
      0.056133845485,
      0.06714236079,
      0.069323441927,
      0.069323441927,
      0.069375883613,
      0.074469370965,
      0.085822811571,
      0.09734322474,
      0.113147192765,
      0.129646562409,
      0.130929753571,
      0.130929753571,
      0.177239286784,
      0.182078514192,
      0.200253195498,
      0.204382397588,
      0.210935173682,
      0.218407436818,
      0.226294385531,
      0.244076946337,
      0.244076946337,
      0.301029995664,
      0.333333333333,
      0.356207187108,
      0.369070246429,
      0.369070246429,
      0.369070246429,
      0.369070246429,
      0.369070246429,
      0.369070246429,
      0.369070246429,
      0.369070246429,
      0.369070246429,
      0.37429503197,
      0.430676558073,
      0.5,
      0.5,
      0.569323441927,
      0.569323441927,
      0.666666666667
    ],
    "tied_queries": 229,
    "worse_queries": 22
  },
  "provenance": {
    "generator_path": "reproduce.py",
    "generator_sha256": "d0a83bffa0c487abafc3f54f1d04c496f16284dc9c566fd917f536b43119a039",
    "outputs": {
      "fig-query-deltas.svg": "b3cacfcc82d84d323ce12948dcf29f123abc8717e46476df43321626e9224ee4"
    },
    "plotting_library": "dependency-free Python SVG generator",
    "study_id": "late-chunking-scifact-matched-content-token-256-v1",
    "verification_command": "python3 reproduce.py --verify"
  },
  "schema_version": 1,
  "title": "Late chunking helps some SciFact queries and hurts others",
  "uncertainty": "The interval uses 20,000 resamples of the 300 query-level paired late-minus-naive differences with seed 20260717. Endpoints are the 2.5th and 97.5th percentiles with linear interpolation at (B - 1) * p. It describes this fixed benchmark, not a population."
}