关于「零幻觉文献核验器:字段、标识符、台账比对与逐条判决」,《多窗口写书防漂移》给出的定量结论是什么?
结论:零幻觉核验分三道。格式检查确认字段齐全、DOI 合语法、书号校验位正确、年份卷期说得通;它拦得住抄写错误,拦不住像样的编造,因为编造者可以先算好校验位、照格式拼出 DOI。台账比对拿作者、年份、标题与出处去对一份逐条核过原件的本地台账,标题相似度不低于 0.90 才可能通过,0.60 到 0.90 之间一律交人判断;加权打分里「只有标题不同」的组合仍有 +11.973 的高分,所以标题必须是闸门而不是加分项。判决分通过、改正、暂扣、拒收四档,交印稿只允许前两档存在。在本章的虚构样本里,借用别人 DOI 的 F6 与嵌合标题的 F5 是格式检查完全放过、只有台账才拦得住的两条;而台账覆盖率越高,暂扣清单越短,每一条暂扣也越值得怀疑。
数据口径:书中未标出处的错误率都是假设参数或合成数据上的测量值;本书不是法律意见,也不推荐任何写作工具或模型。 第八章让四个窗口并发写一本九章合成书,注入五类共 119 个错误:四道闸门拦住 0.723,漏网的 33 个里有 28 个属于机器在结构上看不见的手法,扣除后检出率为 0.945;干净书上术语检测首轮误报 1672 处,人审 48 个词后归零;一份严肃的注入审计每类至少要 75 个错误。
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: 11.973
| A. variant lev bi word | |
| exact | 1.000 1.000 1.000 |
| case and dash | 1.000 1.000 1.000 |
| one typo | 0.981 0.936 0.750 |
| subtitle added | 0.800 0.821 0.700 |
| reordered | 0.500 0.938 1.000 |
| plausible chimera | 0.732 0.618 0.556 |
| same words, new claim | 0.173 0.917 1.000 |
| unrelated | 0.308 0.164 0.000 |
| B. field m u agree w disagree w | |
| title | 0.950 0.001 +9.892 -4.320 |
| first author surname | 0.970 0.010 +6.600 -5.044 |
| year | 0.900 0.050 +4.170 -3.248 |
| venue | 0.920 0.020 +5.524 -3.615 |
| C. pattern total weight | |
| all four agree +26.185 | |
| year off, rest agree +18.767 | |
| title differs, rest agree +11.973 | |
| only title agrees | -2.015 |
| nothing agrees | -16.228 |
查看完整验算程序
# Chapter 5. p54: Title similarity measures, and Fellegi-Sunter field weights, on FICTIONAL titles.
# Normalise: case-fold, drop punctuation, collapse spaces. Three measures against the ledger title:
# lev = 1 - Levenshtein(a, b) / max(len a, len b) (character edits)
# bi = Jaccard of character bigrams (order-tolerant, typo-tolerant)
# word = Jaccard of word sets (order-blind)
# Part B: agreement weights log2(m / u) and disagreement weights log2((1 - m) / (1 - u)) per field.
# The m and u values are ASSUMED for the illustration, not estimated from data.
import math
import re
LEDGER_TITLE = "Ledger-First Citation Checking for Serial Monographs"
VARIANTS = [
("exact", "Ledger-First Citation Checking for Serial Monographs"),
("case and dash", "ledger first citation checking for serial monographs"),
("one typo", "Ledger-First Citaton Checking for Serial Monographs"),
("subtitle added", "Ledger-First Citation Checking for Serial Monographs: A Field Test"),
("reordered", "Citation Checking for Serial Monographs, Ledger-First"),
("plausible chimera", "Ledger-Based Citation Verification for Serial Monographs"),
("same words, new claim", "Serial Monographs for Ledger-First Citation Checking"),
("unrelated", "A Theory of Shelf Rotation in Small Bookshops"),
]
FIELDS = [("title", 0.95, 0.001), ("first author surname", 0.97, 0.01), ("year", 0.90, 0.05),
("venue", 0.92, 0.02)]
def norm(t):
return re.sub(r"\s+", " ", re.sub(r"[^\w\s]", " ", t.casefold())).strip()
def lev(a, b):
prev = list(range(len(b) + 1))
for i, ca in enumerate(a, 1):
cur = [i]
for j, cb in enumerate(b, 1):
cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (ca != cb)))
prev = cur
return prev[-1]
def jac(x, y):
return len(x & y) / len(x | y) if x | y else 1.0
def measures(a, b):
a, b = norm(a), norm(b)
l = 1 - lev(a, b) / max(len(a), len(b))
bi = jac({a[i:i + 2] for i in range(len(a) - 1)}, {b[i:i + 2] for i in range(len(b) - 1)})
return l, bi, jac(set(a.split()), set(b.split()))
def main():
print("A. variant lev bi word")
for name, t in VARIANTS:
l, bi, w = measures(LEDGER_TITLE, t)
print(" {:<22} {:.3f} {:.3f} {:.3f}".format(name, l, bi, w))
print("B. field m u agree w disagree w")
weights = {}
for f, m, u in FIELDS:
weights[f] = (math.log2(m / u), math.log2((1 - m) / (1 - u)))
print(" {:<22} {:.3f} {:.3f} {:>+8.3f} {:>+8.3f}".format(f, m, u, *weights[f]))
patterns = [("all four agree", (1, 1, 1, 1)), ("year off, rest agree", (1, 1, 0, 1)),
("title differs, rest agree", (0, 1, 1, 1)), ("only title agrees", (1, 0, 0, 0)),
("nothing agrees", (0, 0, 0, 0))]
print("C. pattern total weight")
for name, agree in patterns:
tot = math.fsum(weights[f][0] if a else weights[f][1] for (f, _, _), a in zip(FIELDS, agree))
print(" {:<28} {:>+8.3f}".format(name, tot))
if __name__ == "__main__":
main()唯一数据源与引用
范雅福(Fan, Y.). (2026). 《多窗口并行写书防漂移控制、术语跨卷一致性图审计与零幻觉文献核验器》第 5 章「零幻觉文献核验器:字段、标识符、台账比对与逐条判决」(版本 0.1.0). 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-152 · v0.1.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026152,
author = {Fan, Yafu},
title = {Drift Control for Parallel Multi-Window Book Writing, Cross-Volume Terminology Graph Audits and Zero-Hallucination Citation Verifiers},
year = {2026},
publisher = {5D Library},
edition = {0.1.0},
note = {DOI pending}
}