{
 "id": "plate-i",
 "number": "I",
 "title": "The Anatomy",
 "act": {
  "number": "I",
  "title": "The Instrument"
 },
 "corpus_scope": "ICLR 2026 · 410,586 units from 74,380 reviews",
 "question": "What do reviewers actually look at, how do they think, and how do they decide?",
 "page_url": "https://atlas-of-judgment.pages.dev/",
 "figures": [
  {
   "id": "fig-1a",
   "title": "Object → standard → verdict (the sankey)",
   "deck": "71.8% of everything reviewers write is criticism — and two thirds of the praise reaches its verdict through a single standard, merit recognition.",
   "dom_host": "#sankey",
   "claims": [
    {
     "id": "1a-negative-share",
     "statement": "71.8% of all units end negative, and no object escapes — the gentlest-treated object still runs 60% negative, the harshest 88%.",
     "value": {
      "quantity": 0.718,
      "unit": "share_of_units",
      "n": 410586
     },
     "source_island": "viz-data.json",
     "source_path": "triples[]",
     "derivation": "sum(n where verdict=='negative') / sum(n), and per-object min/max of the same ratio",
     "recompute": "scripts/build_unit_viz_data.py",
     "dom_ref": null,
     "verified": "D10 number audit, 2026-08-24",
     "caveat_refs": [
      "unit-recall-bias"
     ]
    },
    {
     "id": "1a-mercy-bottleneck",
     "statement": "Two thirds of all positive units (66.9%) reach the right edge through one standard, merit recognition.",
     "value": {
      "quantity": 0.669,
      "unit": "share_of_positive_units"
     },
     "source_island": "viz-data.json",
     "source_path": "triples[]",
     "derivation": "sum(n where verdict=='positive' and standard=='merit_recognition') / sum(n where verdict=='positive')",
     "recompute": "scripts/build_unit_viz_data.py",
     "dom_ref": "data-sankey=\"mercy-positive\"",
     "verified": "D10 number audit, 2026-08-24",
     "caveat_refs": []
    },
    {
     "id": "1a-largest-current",
     "statement": "The single largest current in the corpus is empirical scope flowing into the design-justification standard (4.8% of everything).",
     "value": {
      "quantity": 0.048,
      "unit": "share_of_all_units"
     },
     "source_island": "viz-data.json",
     "source_path": "triples[]",
     "derivation": "n(object=='empirical_scope', standard=='design_justification') / sum(n)",
     "recompute": "scripts/build_unit_viz_data.py",
     "dom_ref": "data-sankey=\"largest-current\"",
     "verified": "D10 number audit, 2026-08-24",
     "caveat_refs": []
    },
    {
     "id": "1a-novelty-monogamous",
     "statement": "Novelty is the most monogamous object — one standard (its own) carries 67% of its units and two cover 80%.",
     "value": {
      "quantity": 0.67,
      "unit": "share_of_object_units",
      "object": "novelty"
     },
     "source_island": "viz-data.json",
     "source_path": "triples[]",
     "derivation": "top-1 and top-2 standard share of n, grouped by object=='novelty', ranked across all 12 objects",
     "recompute": "scripts/build_unit_viz_data.py",
     "dom_ref": "data-sankey=\"novelty-monogamous\"",
     "verified": "D10 number audit, 2026-08-24",
     "caveat_refs": []
    }
   ]
  },
  {
   "id": "fig-1b",
   "title": "The severity map: where mercy lives",
   "deck": "All 131 readable pairings outside the merit-recognition column condemn more often than not; inside that one column, criticism falls to 14–35% for eleven of the twelve objects.",
   "dom_host": "#mercy",
   "claims": [
    {
     "id": "1b-readable-pairings",
     "statement": "131 readable pairings — 12 × 12 pairings minus the merit-recognition column (12) minus 1 cell under 40 units; every one runs majority-negative.",
     "value": {
      "quantity": 131,
      "unit": "count"
     },
     "source_island": "viz-data.json",
     "source_path": "triples[]",
     "derivation": "12×12 object×standard pairings − merit_recognition column (12) − 1 cell with n<40 = 131",
     "recompute": "scripts/build_unit_viz_data.py",
     "dom_ref": null,
     "verified": "D10 number audit, 2026-08-24",
     "caveat_refs": []
    },
    {
     "id": "1b-harshest-cells",
     "statement": "The harshest cells: related work judged by presentation-as-trust (95.9% negative), clarity judged by the same standard (94.7%), and reproducibility judged by statistical identifiability (94.4%).",
     "value": {
      "quantity": [
       0.959,
       0.947,
       0.944
      ],
      "unit": "share_negative_per_cell"
     },
     "source_island": "viz-data.json",
     "source_path": "triples[]",
     "derivation": "top-3 (object, standard) cells by negative-share, each holding hundreds to thousands of units",
     "recompute": "scripts/build_unit_viz_data.py",
     "dom_ref": "data-mercy=\"harshest-three\"",
     "verified": "D10 number audit, 2026-08-24",
     "caveat_refs": [
      "unit-recall-bias"
     ]
    },
    {
     "id": "1b-rw-no-mercy",
     "statement": "Related work is the worst-treated object even inside the mercy column: 55% negative even under merit recognition, where no other object exceeds 35%.",
     "value": {
      "quantity": 0.55,
      "unit": "share_negative",
      "object": "related_work",
      "standard": "merit_recognition"
     },
     "source_island": "viz-data.json",
     "source_path": "triples[]",
     "derivation": "negative-share of the (related_work, merit_recognition) cell, compared against the same cell for all 12 objects",
     "recompute": "scripts/build_unit_viz_data.py",
     "dom_ref": "data-mercy=\"rw-mercy\"",
     "verified": "D10 number audit, 2026-08-24",
     "caveat_refs": []
    }
   ]
  },
  {
   "id": "fig-1c",
   "title": "The instrument, checked against the venue's own gauges",
   "deck": "Where the atlas's twelve categories and ICLR's own three sub-scores should agree, they do — and the one place they part is itself a finding.",
   "dom_host": "#construct-hm",
   "claims": [
    {
     "id": "1c-clarity-presentation",
     "statement": "A review that criticises clarity sits 0.35 points lower on ICLR's own presentation sub-score at the same overall rating — the sharpest cell of the 12x3 map, four times any other.",
     "value": {
      "quantity": -0.3465,
      "unit": "sub_score_gap_1to4",
      "n": 74380
     },
     "source_island": "construct-data.json",
     "source_path": "matrix.clarity.presentation",
     "derivation": "mean presentation sub-score of reviews with >=1 negative clarity unit minus reviews without, within each overall-rating level, weighted-averaged",
     "recompute": "scripts/build_construct_data.py",
     "dom_ref": null,
     "verified": "checked against shipped JSON, 2026-08-27",
     "caveat_refs": [
      "construct-2026-only"
     ]
    },
    {
     "id": "1c-novelty-contribution",
     "statement": "Novelty criticism dents the contribution sub-score specifically (-0.12, that column's deepest cell).",
     "value": {
      "quantity": -0.1219,
      "unit": "sub_score_gap_1to4",
      "n": 74380
     },
     "source_island": "construct-data.json",
     "source_path": "matrix.novelty.contribution",
     "derivation": "same rating-adjusted gap, novelty row x contribution column",
     "recompute": "scripts/build_construct_data.py",
     "dom_ref": null,
     "verified": "checked against shipped JSON, 2026-08-27",
     "caveat_refs": [
      "construct-2026-only"
     ]
    },
    {
     "id": "1c-theory-leak",
     "statement": "Theory criticism dents presentation (-0.07) three times harder than soundness (-0.02): 'I could not follow the theory' is filed as a writing complaint rather than a rigor complaint.",
     "value": {
      "quantity": -0.0683,
      "unit": "sub_score_gap_1to4",
      "n": 74380
     },
     "source_island": "construct-data.json",
     "source_path": "matrix.theory.presentation vs matrix.theory.soundness",
     "derivation": "-0.0683 (presentation) vs -0.0217 (soundness), same rating-adjusted design",
     "recompute": "scripts/build_construct_data.py",
     "dom_ref": null,
     "verified": "checked against shipped JSON, 2026-08-27",
     "caveat_refs": [
      "construct-2026-only",
      "construct-reviewer-vs-instrument"
     ]
    }
   ]
  }
 ],
 "caveats": [
  {
   "id": "unit-recall-bias",
   "text": "The models that split reviews into units catch explicit criticism more reliably than brief praise — a passing \"well written\" often goes unrecorded — so the negative share is tilted upward. Comparisons between cells stay safe because every object and standard was read the same way.",
   "scope": "all Fig. 1a / 1b percentages",
   "verified_by": "Appendix II"
  },
  {
   "id": "construct-2026-only",
   "text": "Sub-scores exist only in recent review forms; the check runs on ICLR 2026 alone. Association at fixed rating, not causation."
  },
  {
   "id": "construct-reviewer-vs-instrument",
   "text": "The gauge cannot split the reviewer from the instrument: a leak may be the reviewer's filing habit, the model's reading, or both."
  }
 ],
 "links": {
  "recompute_scripts": [
   "scripts/build_unit_viz_data.py"
  ],
  "data": [
   "/api/v1/data/viz-data.json"
  ],
  "related_plates": [
   "plate-ii",
   "plate-xii",
   "plate-xiii"
  ]
 }
}