关于「实证闭环:直答段落模板 A/B、可引用度评分器与预注册检验」,《生成式搜索引用权重》给出的定量结论是什么?

结论:第八章给出直答段落模板的 A/B 检验、小工具「直答段落可引用度评分器」与预注册方案(每组 524 次检查,模拟检验力 0.8085)。书中以标注框区分已发表的证据、可运行的 Python 计算(21 段程序)、原典与作者的假说。

数据口径:书中参数均为示意;全部计算由书中程序生成。 第五章讲可核查性,并用模拟说明从外部推断特征权重的脆弱:漏掉长度时统计数字的权重被高估 0.3252。

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 标准库与结论吻合的打印值: 0.8085, 524

p1 n / arm icc 0.05 icc 0.20
0.141907 2194 3052
0.17524 603 839
0.20253 291 405
simulated power at n524 per arm: 0.8085 (2000 runs)
icc 0.00524 checks per arm = 4.4 weeks at 120 a week
icc 0.05603 checks per arm = 5.0 weeks at 120 a week
icc 0.20839 checks per arm = 7.0 weeks at 120 a week
查看完整验算程序
# Chapter 8. p82: how many checks the pre-registered test needs.
# Two-proportion test, alpha 0.05 two-sided, power 0.80, baseline
# citation rate p0 = 0.11 (close to arm A in p81). Checks are
# clustered: each question goes to m = 4 engines, so the needed count
# grows by the design effect 1 + (m - 1) icc. We then check the
# unclustered answer for p1 = 0.17 by 2000 seeded simulations.
# Rates and icc values are illustrative.
import math
import random

ZA, ZB = 1.959964, 0.841621
P0, M = 0.11, 4


def n_per_arm(p0, p1):
    pbar = (p0 + p1) / 2
    a = ZA * math.sqrt(2 * pbar * (1 - pbar))
    b = ZB * math.sqrt(p0 * (1 - p0) + p1 * (1 - p1))
    return math.ceil((a + b) ** 2 / (p1 - p0) ** 2)


print("%6s %10s %12s %12s" % ("p1", "n / arm", "icc 0.05", "icc 0.20"))
for p1 in (0.14, 0.17, 0.20):
    n = n_per_arm(P0, p1)
    d1, d2 = 1 + (M - 1) * 0.05, 1 + (M - 1) * 0.20
    print("%6.2f %10d %12d %12d" % (p1, n, math.ceil(n * d1),
                                    math.ceil(n * d2)))
rng = random.Random(20261113)
n = n_per_arm(P0, 0.17)
hits = 0
SIMS = 2000
for _ in range(SIMS):
    ka = len([1 for _ in range(n) if rng.random() < P0])
    kb = len([1 for _ in range(n) if rng.random() < 0.17])
    pp = (ka + kb) / (2 * n)
    se = math.sqrt(pp * (1 - pp) * 2 / n)
    if se > 0 and (kb - ka) / n / se > ZA:
        hits += 1
print("simulated power at n = %d per arm: %.4f (%d runs)" % (
    n, hits / SIMS, SIMS))
for icc in (0.0, 0.05, 0.20):
    need = math.ceil(n * (1 + (M - 1) * icc))
    print("icc %.2f: %d checks per arm = %.1f weeks at 120 a week" % (
        icc, need, need / 120))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《生成式搜索引用权重几何学、问答引擎检索重排特征与高被引双语直答页段落拓扑母典》第 8 章「实证闭环:直答段落模板 A/B、可引用度评分器与预注册检验」(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-260 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026260,
  author    = {Fan, Yafu},
  title     = {The Geometry of Citation Weight in Generative Search: Retrieval and Reranking Features and the Paragraph Topology of Highly Cited Bilingual Answer Pages},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论