{
  "id": "plate-vi",
  "number": "VI",
  "title": "The Early Verdict",
  "act": {
    "number": "II",
    "title": "One Hand"
  },
  "corpus_scope": "ICLR 2026 · 74,380 rated reviews",
  "question": "How much of a reviewer's own final rating can be read off the review's own text, how quickly does reading more of it stop helping, and is the tally that predicts it weighted or flat?",
  "page_url": "https://atlas-of-judgment.pages.dev/",
  "figures": [
    {
      "id": "fig-6a",
      "title": "Verdict readability against units read",
      "deck": "The rating is a weighted tally: a novelty negative counts ×2.4 the average, a robustness negative ×0.4 — yet the count alone does almost all the ranking, and nothing after the ninth unit changes it.",
      "dom_host": "#commit-fig",
      "claims": [
        {
          "id": "6a-verdict-contribution",
          "statement": "Verdicts alone buy +0.260 of the whole review's +0.279 AUC climb toward predicting the reviewer's own rating.",
          "value": {
            "quantity": 0.26,
            "unit": "auc_contribution"
          },
          "source_island": "commit-data.json",
          "source_path": "auc / auc_full",
          "derivation": "AUC gained by adding verdict features vs. the whole review's total AUC gain over chance (0.279)",
          "recompute": "scripts/build_commit_data.py",
          "dom_ref": null,
          "verified": "D10 number audit, 2026-08-24",
          "caveat_refs": [
            "internal-consistency-not-prediction"
          ]
        },
        {
          "id": "6a-topic-contribution",
          "statement": "Topics add only +0.019 AUC on top of verdicts across all reviews; in the long-review subset the topic contribution is -0.006 — noise.",
          "value": {
            "quantity": [
              0.019,
              -0.006
            ],
            "unit": "auc_gain_from_topics_all_vs_long"
          },
          "source_island": "commit-data.json",
          "source_path": "auc_full / auc_long_full",
          "derivation": "(auc_full - auc) for all reviews vs. (auc_long_full - auc_long) for the 10+-unit long-review subset",
          "recompute": "scripts/build_commit_data.py",
          "dom_ref": null,
          "verified": "D10 number audit, 2026-08-24",
          "caveat_refs": [
            "internal-consistency-not-prediction"
          ]
        },
        {
          "id": "6a-novelty-weight",
          "statement": "One novelty negative counts as 2.4 average negatives — six robustness negatives.",
          "value": {
            "quantity": 2.4,
            "unit": "relative_weight_vs_average_negative"
          },
          "source_island": "commit-data.json",
          "source_path": "tally",
          "derivation": "tally.coef['novelty'] / mean(tally.coef.*), a logistic fit on whole-review per-topic negative counts",
          "recompute": "scripts/build_commit_data.py",
          "dom_ref": "data-commit=\"weights\"",
          "verified": "D10 number audit, 2026-08-24",
          "caveat_refs": [
            "association-not-causation"
          ]
        },
        {
          "id": "6a-tariff-correlation",
          "statement": "The ordering of these per-topic weights reproduces the within-paper tariff of Plate XIV almost exactly (Spearman ρ = 0.90).",
          "value": {
            "quantity": 0.9,
            "unit": "spearman_rho"
          },
          "source_island": "commit-data.json",
          "source_path": "tally",
          "derivation": "Spearman rank correlation between tally.coef ordering and Plate XIV's jurisprudence-data.json tariff.*.coef ordering",
          "recompute": "scripts/build_commit_data.py",
          "dom_ref": null,
          "verified": "D10 number audit, 2026-08-24",
          "caveat_refs": [
            "association-not-causation"
          ]
        },
        {
          "id": "6a-first-unit-auc",
          "statement": "One unit reaches 0.69 AUC, two thirds of the way to the 0.78 ceiling.",
          "value": {
            "quantity": [
              0.69,
              0.78
            ],
            "unit": "auc"
          },
          "source_island": "commit-data.json",
          "source_path": "auc",
          "derivation": "auc[k=1] compared to the whole-review ceiling auc[k=max]",
          "recompute": "scripts/build_commit_data.py",
          "dom_ref": "data-commit=\"first-unit\"",
          "verified": "D10 number audit, 2026-08-24",
          "caveat_refs": [
            "internal-consistency-not-prediction"
          ]
        },
        {
          "id": "6a-words-floor",
          "statement": "The words themselves reach 0.822 AUC (TF-IDF over the full text, a floor not a ceiling on what the text carries) — more than the coarse verdict-plus-topic profile (+0.043).",
          "value": {
            "quantity": [
              0.822,
              0.043
            ],
            "unit": "auc_and_auc_gain_over_profile"
          },
          "source_island": "commit-data.json",
          "source_path": "text_auc",
          "derivation": "text_auc (TF-IDF classifier AUC) compared to auc_full (verdict+topic profile AUC)",
          "recompute": "scripts/build_commit_data.py",
          "dom_ref": null,
          "verified": "D10 number audit, 2026-08-24",
          "caveat_refs": [
            "internal-consistency-not-prediction"
          ]
        }
      ]
    }
  ],
  "caveats": [
    {
      "id": "internal-consistency-not-prediction",
      "text": "The target is the same reviewer's own rating, so this measures the internal consistency of a document with its own summary number, not a prediction of anything external — and it is a statement about the text, not about when the reviewer decided.",
      "scope": "all Fig. 6a AUC findings",
      "verified_by": null
    },
    {
      "id": "association-not-causation",
      "text": "Panel ⑤'s weights are descriptive regression coefficients on correlated counts, not causal prices; their agreement with Plate XIV's independently-designed tariff is what earns them the reading.",
      "scope": "Fig. 6a novelty-weight and tariff-correlation findings",
      "verified_by": null
    }
  ],
  "links": {
    "recompute_scripts": [
      "scripts/build_commit_data.py",
      "scripts/build_score_data.py"
    ],
    "data": [
      "/api/v1/data/commit-data.json",
      "/api/v1/data/score-data.json"
    ],
    "related_plates": [
      "plate-iv",
      "plate-v",
      "plate-xiv",
      "plate-xxv"
    ]
  }
}