关于「多模型裁决去相关的实证闭环:流程、相关投票准确率计算器与预注册证伪方案」,《相关误差下的多模型投票》给出的定量结论是什么?

结论:第八章给出多模型裁决去相关的实证闭环、蒸馏小工具「相关投票准确率计算器」与预注册证伪方案。书中以标注框区分已发表的证据、可运行的 Python 计算(19 段程序)、原典与作者的假说。

数据口径:书中参数为示意;全部计算由书中程序生成;本书是非个人化的经营分析,不构成经营建议。 书中以标注框区分已发表的证据、可运行的 Python 计算(19 段程序)、原典与作者的假说。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库

target accuracy 0.80
5 resamples of one model
  final 0.7316  (independent benchmark 0.8624)
  ceiling 0.7366  n_eff 1.47  cap 1.7
  verdict DECORRELATE
5 different models
  final 0.7533  (independent benchmark 0.8095)
  ceiling 0.8257  n_eff 3.12  cap 6.7
  verdict ADD-VOTES  (n = 21)
5 different models + check
  final 0.8983  (independent benchmark 0.8095)
  ceiling 0.9273  n_eff 3.12  cap 6.7
  verdict PASS
查看完整验算程序
# Chapter 8. p81: the audit of a multi-model verdict, as in the tool.
# Model: a pool of n votes with measured mean accuracy p and error
# correlation rho (exchangeable Beta-binomial, p21), optionally a
# standard-library check covering a share cov of questions with
# accuracy pc that decides where it applies (p42). Reported: the
# accuracy at n, the independent benchmark, the ceiling, n_eff and its
# cap 1/rho, the calls a target needs, and a verdict:
#   PASS        the design already reaches the target
#   ADD-VOTES   more votes of the same pool reach it (n given)
#   DECORRELATE the ceiling lies below the target: more of the same
#               pool can never reach it; change prompts, makers, or
#               add a deterministic check
# The three cases are illustrative.
import math



def lbeta(a, b):
    return math.lgamma(a) + math.lgamma(b) - math.lgamma(a + b)


def ab(p, rho):
    s = 1 / rho - 1
    return p * s, (1 - p) * s


def bb_major(n, p, rho):
    if rho == 0:
        lp, lq = math.log(p), math.log(1 - p)
        terms = [math.exp(math.lgamma(n + 1) - math.lgamma(k + 1)
                          - math.lgamma(n - k + 1) + k * lp
                          + (n - k) * lq)
                 for k in range(n // 2 + 1, n + 1)]
        return math.fsum(terms)
    a, b = ab(p, rho)
    terms = [math.exp(math.lgamma(n + 1) - math.lgamma(k + 1)
                      - math.lgamma(n - k + 1) + lbeta(k + a, n - k + b)
                      - lbeta(a, b))
             for k in range(n // 2 + 1, n + 1)]
    return math.fsum(terms)


def betacf(a, b, x):
    tiny = 1e-300
    c, d = 1.0, 1.0 - (a + b) * x / (a + 1)
    d = 1.0 / (d if abs(d) > tiny else tiny)
    h = d
    for m in range(1, 400):
        m2 = 2 * m
        for num in (m * (b - m) * x / ((a + m2 - 1) * (a + m2)),
                    -(a + m) * (a + b + m) * x / ((a + m2) * (a + m2 + 1))):
            d = 1.0 + num * d
            d = 1.0 / (d if abs(d) > tiny else tiny)
            c = 1.0 + num / c
            c = c if abs(c) > tiny else tiny
            h *= d * c
        if abs(d * c - 1.0) < 1e-15:
            break
    return h


def betai(a, b, x):
    if x <= 0:
        return 0.0
    if x >= 1:
        return 1.0
    lead = math.exp(a * math.log(x) + b * math.log(1 - x) - lbeta(a, b))
    if x < (a + 1) / (a + b + 2):
        return lead * betacf(a, b, x) / a
    return 1.0 - lead * betacf(b, a, 1 - x) / b


def limit(p, rho):
    if rho == 0:
        return 1.0 if p > 0.5 else 0.0
    a, b = ab(p, rho)
    return 1.0 - betai(a, b, 0.5)


def calls(p, rho, target, cap=3001):
    if limit(p, rho) <= target:
        return None
    n = 1
    while bb_major(n, p, rho) < target - 1e-12 and n < cap:
        n += 2
    return n if n < cap else None


def audit(n, p, rho, target, cov=0.0, pc=0.995):
    acc = bb_major(n, p, rho)
    final = cov * pc + (1 - cov) * acc
    ceil = limit(p, rho)
    top = cov * pc + (1 - cov) * ceil
    neff = n / (1 + (n - 1) * rho)
    if final >= target:
        verdict = "PASS"
    elif top <= target:
        verdict = "DECORRELATE"
    else:
        verdict = "ADD-VOTES"
    need = calls(p, rho, (target - cov * pc) / (1 - cov)) \
        if verdict == "ADD-VOTES" else None
    return {"acc": acc, "indep": bb_major(n, p, 0.0), "final": final,
            "ceiling": top, "n_eff": neff,
            "cap": 1 / rho if rho > 0 else float("inf"),
            "verdict": verdict, "need": need}


if __name__ == "__main__":
    cases = (("5 resamples of one model", 5, 0.72, 0.6, 0.0),
             ("5 different models", 5, 0.68, 0.15, 0.0),
             ("5 different models + check", 5, 0.68, 0.15, 0.6))
    target = 0.8
    print("target accuracy {:.2f}".format(target))
    for name, n, p, rho, cov in cases:
        r = audit(n, p, rho, target, cov)
        print(name)
        print("  final {:.4f}  (independent benchmark {:.4f})".format(
            r["final"], r["indep"]))
        print("  ceiling {:.4f}  n_eff {:.2f}  cap {:.1f}".format(
            r["ceiling"], r["n_eff"], r["cap"]))
        tail = "" if r["need"] is None else \
            "  (n = {})".format(r["need"])
        print("  verdict {}{}".format(r["verdict"], tail))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《孔多塞陪审团定理与相关误差、多模型投票何时越投越准何时同源共错母典》第 8 章「多模型裁决去相关的实证闭环:流程、相关投票准确率计算器与预注册证伪方案」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-309 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026309,
  author    = {Fan, Yafu},
  title     = {Condorcet's Jury Theorem with Correlated Errors: When Multi-Model Voting Gets More Accurate and When Shared Origins Make Everyone Wrong Together},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论