关于「隐式指纹:码长下界与改写衰减」,《合同暗坑与资产主权》给出的定量结论是什么?

结论:「作品发出去就无从追查」是错的:做过标记的副本带着来源的痕迹。「百分之百可追溯」也是错的:塔尔多什码长随合谋人数按平方增长,原始常数下十二种组合全都装不进一本书 700 位的预算;改写比例超过 0.8 时,400 位标记的溯源成功率跌到 0.28。指纹是有物理上界的工具,规则要按批次、按位数、按改写强度来写。

数据口径:合同机检确有扎实证据:法律专家在 510 份商业合同里标注了 41 类审阅要点,在线服务条款里的八类潜在不公平条款可被机器标出;但机检是召回优先的初筛,在本书的合成标注集上,改写过的风险条款只有 0.380 被抓到,漏报代价越高,律师越要通读全文。 书中参数为示意,全部计算由书中程序生成。

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: 0.28, 0.8, 400

accuse line z > 4.42 (alpha 0.001 over 200 copies)
Krz theoryz meanmax innocentfound
4000.020.0020.004.301.00
4000.216.0016.024.201.00
4000.412.0012.024.101.00
4000.68.008.004.301.00
4000.76.006.083.900.99
4000.84.004.043.700.28
4000.92.002.034.100.03
1000.82.001.763.600.00
1,6000.88.007.913.751.00
查看完整验算程序
# Chapter 4. p43: a text fingerprint fading as a leaked copy is reworded
# Each of n = 200 copies of our own manuscript carries K binary slots
# (synonym or punctuation choices) set by a keyed hash of copy and slot.
# A leaked copy has each slot reworded with probability r; a reworded
# slot becomes a fair coin. For each copy, z = (matches - K/2) /
# sqrt(K/4), so E[z of the true copy] = sqrt(K) (1 - r). Accuse the top
# copy only if its z clears the Bonferroni line for n copies at alpha =
# 0.001. Fixed seeds; demo key string; all values illustrative.
import hashlib
import math
import random
from statistics import NormalDist

N_COPIES, ALPHA, TRIALS = 200, 0.001, 100
LINE = NormalDist().inv_cdf(1 - ALPHA / N_COPIES)


def patterns(k):
    pats = []
    for j in range(N_COPIES):
        v = 0
        for s in range(k):
            h = hashlib.sha256("demo-key|{}|{}".format(j, s).encode())
            v = (v << 1) | (h.digest()[0] & 1)
        pats.append(v)
    return pats


def run(k, r, pats, rng):
    found = 0
    zs, worst = [], []
    for _ in range(TRIALS):
        src = rng.randrange(N_COPIES)
        leak = pats[src]
        for s in range(k):
            if rng.random() < r:
                bit = rng.randint(0, 1)
                leak = (leak & ~(1 << s)) | (bit << s)
        z = []
        for j in range(N_COPIES):
            match = k - bin(pats[j] ^ leak).count("1")
            z.append((match - k / 2) / math.sqrt(k / 4))
        top = max(range(N_COPIES), key=lambda j: z[j])
        if top == src and z[top] > LINE:
            found += 1
        zs.append(z[src])
        worst.append(max(z[j] for j in range(N_COPIES) if j != src))
    return (math.fsum(zs) / TRIALS, max(worst), found / TRIALS)


rng = random.Random(38343)
print("accuse line z > {:.2f} (alpha {} over {} copies)".format(
    LINE, ALPHA, N_COPIES))
print("")
print("{:>6}{:>6}{:>10}{:>10}{:>14}{:>10}".format(
    "K", "r", "z theory", "z mean", "max innocent", "found"))
pats = patterns(400)
for r in (0.0, 0.2, 0.4, 0.6, 0.7, 0.8, 0.9):
    zm, wi, f = run(400, r, pats, rng)
    print("{:>6}{:>6.1f}{:>10.2f}{:>10.2f}{:>14.2f}{:>10.2f}".format(
        400, r, 20 * (1 - r), zm, wi, f))
print("")
for k in (100, 1600):
    pk = patterns(k)
    zm, wi, f = run(k, 0.8, pk, rng)
    print("{:>6,}{:>6.1f}{:>10.2f}{:>10.2f}{:>14.2f}{:>10.2f}".format(
        k, 0.8, math.sqrt(k) * 0.2, zm, wi, f))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《人机联防资产与知识产权护城河、契约暗坑机检与全方位防侵蚀母典》第 4 章「隐式指纹:码长下界与改写衰减」(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-383 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026383,
  author    = {Fan, Yafu},
  title     = {Human-AI Joint Asset and IP Protection, Automated Contract-Loophole Auditing and Full-Spectrum Anti-Erosion Architecture},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论