{
  "artifact": "Uncertainty and self-critique analysis of the published Study 1 run",
  "source_data": "swarm-study1-per-item.csv (published per-item predictions)",
  "method": "Paired bootstrap over items (20000 resamples, seed 20260701), swarm minus baseline; exact McNemar for accuracy; analytic binomial illustration for consensus granularity. Stdlib only; no new data.",
  "n_items": 66,
  "n_incorrect_items": 12,
  "headline_with_uncertainty": {
    "note": "Point estimates match the published findings JSON; the CIs are new.",
    "swarm": {
      "accuracy": 0.8182,
      "mean_confidence": 0.9293,
      "overconfidence": 0.1111,
      "ece": 0.1111,
      "auc": 0.7407
    },
    "baseline": {
      "accuracy": 0.8182,
      "mean_confidence": 0.8702,
      "overconfidence": 0.052,
      "ece": 0.109,
      "auc": 0.8295
    },
    "difference_swarm_minus_baseline": {
      "accuracy": {
        "point": 0.0,
        "ci95": [
          -0.0455,
          0.0455
        ],
        "excludes_zero": false,
        "resamples_used": 20000
      },
      "mean_confidence": {
        "point": 0.0591,
        "ci95": [
          0.0162,
          0.1058
        ],
        "excludes_zero": true,
        "resamples_used": 20000
      },
      "overconfidence": {
        "point": 0.0591,
        "ci95": [
          0.0115,
          0.1108
        ],
        "excludes_zero": true,
        "resamples_used": 20000
      },
      "ece": {
        "point": 0.0021,
        "ci95": [
          -0.0701,
          0.0607
        ],
        "excludes_zero": false,
        "resamples_used": 20000
      },
      "auc": {
        "point": -0.0887,
        "ci95": [
          -0.2765,
          0.083
        ],
        "excludes_zero": false,
        "resamples_used": 20000
      }
    },
    "confidence_gap_paired_t": {
      "mean_diff": 0.0591,
      "t": 2.56,
      "df": 65
    },
    "accuracy_mcnemar": {
      "discordant_items": 2,
      "p_two_sided": 1.0
    }
  },
  "preregistered_difficulty_prediction": {
    "prediction": "Pre-registered Study 1 hypothesis: convergence-confidence decouples from accuracy as difficulty rises (ECE grows on the hard tail). Verdict in this run: NOT OBSERVED in this slice \u2014 reported as a null, not reframed.",
    "by_difficulty": {
      "hard": {
        "n": 12,
        "swarm": {
          "accuracy": 0.8333,
          "mean_confidence": 0.9167,
          "overconfidence": 0.0833,
          "ece": 0.0833,
          "auc": 0.95
        },
        "baseline": {
          "accuracy": 0.75,
          "mean_confidence": 0.8026,
          "overconfidence": 0.0526,
          "ece": 0.132,
          "auc": 0.9259
        }
      },
      "medium": {
        "n": 54,
        "swarm": {
          "accuracy": 0.8148,
          "mean_confidence": 0.9321,
          "overconfidence": 0.1173,
          "ece": 0.1173,
          "auc": 0.6989
        },
        "baseline": {
          "accuracy": 0.8333,
          "mean_confidence": 0.8852,
          "overconfidence": 0.0519,
          "ece": 0.1264,
          "auc": 0.7938
        }
      }
    },
    "caveat": "The battery has no easy stratum (54 medium / 12 hard) and difficulty comes from loader heuristics, so this run is a weak test of the prediction by construction. Battery v2 fixes the stratification before the next paid run."
  },
  "contact_axis_preview": {
    "note": "Overconfidence split by the CSV's tool_verifiable flag. Read with care: the flag counts retrieval as a tool, so it includes the trivia items alongside arithmetic. The published correction's per-source analysis is the sharper cut \u2014 arithmetic is close to honest in both arms and the overconfidence concentrates on trivia-recall (see swarm-study1-correction-analysis.json). Both views agree on the pattern Study 2's contact factor tests: confidence runs furthest ahead where contact was available but unused.",
    "by_tool_verifiability": {
      "tool_verifiable": {
        "n": 41,
        "swarm_overconfidence": 0.2033,
        "baseline_overconfidence": 0.125
      },
      "not_tool_verifiable": {
        "n": 25,
        "swarm_overconfidence": -0.04,
        "baseline_overconfidence": -0.0676
      }
    }
  },
  "granularity_artifact": {
    "explanation": "Swarm confidence is a modal share over 3 agents (possible values 1/3, 2/3, 1); baseline confidence is a modal share over about 7 resamples (varies per item under the matched-token budget; mean 7.06, range 6-8; n=7 is the representative support used here). Coarser supports are mechanically higher. In an idealised iid two-answer world at this run's pooled accuracy, the expected modal share differs between n=3 and n=7 by the amount below with no debate psychology at all. This is an illustration of pure granularity under an independence idealisation (real swarm agents are correlated), not a decomposition of the observed gap; the clean fix is Study 1b's granularity-matched baseline.",
    "pooled_accuracy": 0.8182,
    "expected_modal_share_n3": 0.8512,
    "expected_modal_share_n7": 0.8225,
    "expected_gap_from_granularity_alone": 0.0287,
    "observed_confidence_gap": 0.0591
  },
  "swarm_confidence_distribution": {
    "1": 53,
    "0.6667": 12,
    "0.3333": 1
  },
  "what_this_changes": "The confidence/overconfidence gap survives its confidence interval but a material share of it is plausibly measurement granularity rather than debate psychology; the AUC difference does not survive (CI includes zero with only 12 incorrect items); the accuracy tie is exact (2 discordant items, p=1.0). The published qualitative claim \u2014 debate bought confidence, not accuracy \u2014 stands, but its quantitative size is unsettled pending Study 1b's granularity-matched re-run."
}