{
  "id": "appendix-iii",
  "number": "App. III",
  "title": "The Null Cabinet",
  "act": {
    "number": null,
    "title": "Appendices"
  },
  "corpus_scope": null,
  "question": "Which analyses looked like findings but didn't survive a harder look?",
  "page_url": "https://atlas-of-judgment.pages.dev/",
  "figures": [
    {
      "id": "fig-mirage",
      "title": "The archetype mirage, drawn: real roses against forged ones",
      "deck": "The forgery blooms: five specialists at ×2.0–3.0 emerge from pure multinomial noise, and the clustering explains barely more of the real reviewers (35.8%) than of the noise (31.0%).",
      "dom_host": "#mirage-fig",
      "claims": [
        {
          "id": "mirage-explained-variance",
          "statement": "The clustering explains barely more of the real reviewers (35.8%) than of the noise (31.0%).",
          "value": {
            "quantity": [
              0.358,
              0.31
            ],
            "unit": "share_between_profile_variance_explained",
            "labels": [
              "real",
              "synthetic_null"
            ]
          },
          "source_island": "mirage-data.json",
          "source_path": "ev_real, ev_null",
          "derivation": "share of between-profile variance the 5-cluster k-means partition explains: 0.3578 (35.8%) for the 98,513 real reviewer profiles, 0.3102 (31.0%) for the synthetic multinomial-noise profiles",
          "recompute": "scripts/build_mirage_data.py",
          "dom_ref": null,
          "verified": "checked against shipped JSON, 2026-08-25",
          "caveat_refs": [
            "mirage-quantified-not-total"
          ]
        },
        {
          "id": "mirage-real-clusters",
          "statement": "Top row: the five clusters k-means (k = 5, seed 7) finds among 98,513 real reviewer-paper profiles (official reviewers, ≥5 units, 2018–2026).",
          "value": {
            "quantity": 98513,
            "unit": "count_reviewer_paper_profiles"
          },
          "source_island": "mirage-data.json",
          "source_path": "real[]",
          "derivation": "n_profiles = 98,513 real reviewer-paper profiles (official reviewers with ≥5 units, min_units = 5), clustered k-means with k=5, seed 7",
          "recompute": "scripts/build_mirage_data.py",
          "dom_ref": "data-mirage=\"real\"",
          "verified": "claim recomputed from source island, 2026-08-25",
          "caveat_refs": [
            "small-profile-noise"
          ]
        },
        {
          "id": "mirage-null-clusters",
          "statement": "Bottom row: the five clusters the same pipeline finds in synthetic profiles drawn from one shared field mix — profiles that by construction contain no reviewer types at all.",
          "value": {
            "quantity": null,
            "unit": "synthetic_null_profiles_same_pipeline"
          },
          "source_island": "mirage-data.json",
          "source_path": "null[]",
          "derivation": "same k-means (k=5) partition applied to synthetic profiles drawn from one shared multinomial field mix, null seed 46, at each reviewer's own unit volume",
          "recompute": "scripts/build_mirage_data.py",
          "dom_ref": "data-mirage=\"null\"",
          "verified": "claim recomputed from source island, 2026-08-25",
          "caveat_refs": [
            "small-profile-noise"
          ]
        },
        {
          "id": "mirage-overdispersion",
          "statement": "The mirage is quantified, not total: the real rows sit 1.15× above the multinomial variance floor at the median standard (1.58× at the most dispersed), and the clustering explains 4.8pp more of the real profiles than of the forged ones, a gap that widens with profile length — a faint true signature, far below what the roses' confidence suggests.",
          "value": {
            "quantity": [
              1.15,
              1.58,
              0.048
            ],
            "unit": "overdispersion_ratio_and_pp_gap"
          },
          "source_island": "mirage-data.json",
          "source_path": "overdisp_median, overdisp_max, ev_real, ev_null",
          "derivation": "overdisp_median = 1.15, overdisp_max = 1.584 ≈ 1.58, and ev_real − ev_null = 0.3578 − 0.3102 = 0.0476 ≈ 4.8pp",
          "recompute": "scripts/build_mirage_data.py",
          "dom_ref": null,
          "verified": "claim recomputed from source island, 2026-08-25",
          "caveat_refs": [
            "mirage-quantified-not-total"
          ]
        },
        {
          "id": "cabinet-mirage-plate-removed",
          "statement": "A rose-diagram plate once claimed five reviewer archetypes — Architect, Stress-Tester, Advocate, Gatekeeper, Auditor — from k-means over 98,513 per-reviewer standard profiles, each rose a specialization of ×2.4–3.1 the field. The plate was removed on 2026-08-21.",
          "value": {
            "quantity": [
              2.4,
              3.1
            ],
            "unit": "specialization_ratio_range",
            "plate_removed_date": "2026-08-21"
          },
          "source_island": "mirage-data.json",
          "source_path": "real[].ratio",
          "derivation": "each real cluster's own leading (top) standard ratio, ranging from ×2.43 to ×3.1 across the five clusters — the same real[] clusters this appendix's deflating test reruns",
          "recompute": "scripts/build_mirage_data.py",
          "dom_ref": null,
          "verified": "claim recomputed from source island, 2026-08-25",
          "caveat_refs": [
            "mirage-quantified-not-total",
            "small-profile-noise"
          ]
        },
        {
          "id": "mirage-null-specialists",
          "statement": "The forgery blooms: five specialists at ×2.0–3.0 emerge from pure multinomial noise.",
          "value": {
            "quantity": [
              2.0,
              3.0
            ],
            "unit": "leading_standard_ratio_range",
            "labels": [
              "synthetic_null"
            ]
          },
          "source_island": "mirage-data.json",
          "source_path": "null[].ratio",
          "derivation": "each synthetic-null cluster's own leading (top) standard ratio, ranging from ×2.03 to ×2.96 across the five clusters — despite the null profiles containing no reviewer types by construction",
          "recompute": "scripts/build_mirage_data.py",
          "dom_ref": null,
          "verified": "claim recomputed from source island, 2026-08-25",
          "caveat_refs": [
            "mirage-quantified-not-total",
            "small-profile-noise"
          ]
        }
      ]
    },
    {
      "id": "fig-15b",
      "title": "Acceptance rate by breadth of criticism (relocated to the Null Cabinet, 2026-08-25 — expected by construction)",
      "deck": "Each additional object criticised comes with about five points less acceptance.",
      "dom_host": "#gauntlet",
      "claims": [
        {
          "id": "15b-staircase",
          "statement": "Acceptance falls steeply and then almost linearly — 69% at two objects, 57% at three, then roughly five points for each additional object, down to 27% at nine or more.",
          "value": {
            "quantity": [
              0.6923,
              0.5663,
              0.2738
            ],
            "unit": "acceptance_rate_by_breadth"
          },
          "source_island": "panel-data.json",
          "source_path": "gauntlet",
          "derivation": "acceptance rate of 2026 decided papers grouped by count of distinct objects criticised (breadth 2, 3, 9-or-more), papers with fewer than 30 at a breadth dropped",
          "recompute": "scripts/build_panel_data.py",
          "dom_ref": "data-gaunt=\"staircase\"",
          "verified": "claim recomputed from source island, 2026-08-25",
          "caveat_refs": [
            "breadth-entangled"
          ]
        },
        {
          "id": "15b-modal-seven",
          "statement": "The most common paper is criticised on seven of the twelve objects.",
          "value": {
            "quantity": 7,
            "unit": "modal_breadth_of_objects_criticised",
            "n": 3459
          },
          "source_island": "panel-data.json",
          "source_path": "gauntlet",
          "derivation": "breadth 7 holds the largest paper count (3,459) among the gauntlet buckets",
          "recompute": "scripts/build_panel_data.py",
          "dom_ref": "data-gaunt=\"modal-seven\"",
          "verified": "claim recomputed from source island, 2026-08-25",
          "caveat_refs": []
        }
      ]
    }
  ],
  "caveats": [
    {
      "id": "mirage-quantified-not-total",
      "text": "The mirage is quantified, not total: the real rows sit 1.15× above the multinomial variance floor at the median standard (1.58× at the most dispersed), and the clustering explains 4.8pp more of the real profiles than of the forged ones, a gap that widens with profile length — a faint true signature, far below what the roses' confidence suggests.",
      "scope": "all Appendix III mirage claims",
      "verified_by": null
    },
    {
      "id": "small-profile-noise",
      "text": "A median profile holds six units, and at that grain k-means carves sampling noise into confident-looking specialists. This is the test that deleted a plate.",
      "scope": "all Appendix III mirage claims",
      "verified_by": null
    },
    {
      "id": "breadth-entangled",
      "text": "Association only, and here the two quantities are entangled by construction — a weaker paper gives reviewers more to criticise, so breadth of criticism is partly a measurement of the paper rather than a treatment applied to it.",
      "scope": "Fig. 15b breadth figures",
      "verified_by": null
    }
  ],
  "links": {
    "recompute_scripts": [
      "scripts/build_mirage_data.py"
    ],
    "data": [
      "/api/v1/data/mirage-data.json"
    ],
    "related_plates": [
      "plate-xxvi",
      "plate-iv",
      "plate-xix",
      "plate-xxi",
      "plate-xxiii"
    ]
  }
}