关于「检出率的真相:哈希只看得见被锚定的东西」,《衍生品的出处与保真》给出的定量结论是什么?
结论:第三章把检出率分解为锚定覆盖率与结构检出,并说明核对必须对母书而不对上一层;锚点比对在一组句对上准确率为 1.000。
数据口径:书中参数为示意,代价用记账点数而非价格;全部计算由书中程序生成。 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。
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。 与结论吻合的打印值: 1.000
| detection rate D = c + (1 - c) s | |
| c s | 0.0 s=0.3 s=0.6 |
| 0.25 | 0.250 0.475 0.700 |
| 0.50 | 0.500 0.650 0.800 |
| 0.75 | 0.750 0.825 0.900 |
| 0.90 | 0.900 0.930 0.960 |
| 1.00 | 1.000 1.000 1.000 |
| drifted facts after 3 layers | 5.71 of 40 |
| c | 0.50 s=0.3: undetected 1.997 |
| c | 0.90 s=0.6: undetected 0.228 |
| c | 1.00 s=0.0: undetected 0.000 |
| review h | 0.6: wrong facts left after layer k (of 40) |
| k vs-parent vs-master | |
| 1 | 0.800 0.800 |
| 2 | 1.584 1.104 |
| 3 | 2.352 1.220 |
| 4 | 3.105 1.263 |
| 5 | 3.843 1.280 |
| 6 | 4.566 1.286 |
查看完整验算程序
# Chapter 3. p31: a hash only sees what was anchored.
# Detection rate of a drifted fact: D = c + (1 - c) * s, where c is the
# share of facts anchored by a provenance hash (any change to them is
# caught) and s is the share of drift in unanchored facts that a
# structure check still catches. Then: 40 facts, per-layer fidelity
# 0.95, three layers. Last, human review (catch chance 0.6 per look)
# done against the parent versus against the master.
# All parameters are illustrative.
N, P, K = 40, 0.95, 3
CS = (0.25, 0.50, 0.75, 0.90, 1.00)
SS = (0.0, 0.3, 0.6)
print("detection rate D = c + (1 - c) s")
print("c " + "".join(" s=%.1f" % s for s in SS))
for c in CS:
print("%.2f " % c + "".join(" %5.3f" % (c + (1 - c) * s)
for s in SS))
drift = N * (1 - P ** K)
print("drifted facts after %d layers: %.2f of %d" % (K, drift, N))
for c, s in ((0.50, 0.3), (0.90, 0.6), (1.00, 0.0)):
d = c + (1 - c) * s
print(" c=%.2f s=%.1f: undetected %.3f" % (c, s, drift * (1 - d)))
H, E = 0.6, 1 - P
print("review h=%.1f: wrong facts left after layer k (of %d)" % (H, N))
print("k vs-parent vs-master")
wp = wm = 0.0
for k in range(1, 7):
wp = wp + (1 - wp) * E * (1 - H)
wm = (wm + (1 - wm) * E) * (1 - H)
print("%d %9.3f %9.3f" % (k, N * wp, N * wm))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《母书多模态衍生编译与版权链、单部专著到播客投影片题库的结构同态流水线母典》第 3 章「检出率的真相:哈希只看得见被锚定的东西」(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-268 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026268,
author = {Fan, Yafu},
title = {Multimodal Derivative Compilation and Provenance Chains: A Structure-Preserving Pipeline from One Monograph to Podcasts, Slides and Quiz Banks},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}