What does Contract Pitfalls and Asset Sovereignty conclude, in numbers, about “Hidden Fingerprints: Floors on Code Length and the Decay of Rewording”?

Answer: "Once a work is out, there is no way to trace it" is wrong: a marked copy carries the trace of its source. "Everything can be traced" is wrong too. Tardos code length grows with the square of the number of colluders; with the original constant, none of the twelve combinations fits inside a book's budget of 700 positions; and once rewording passes 0.8 of the text, a fingerprint of 400 positions traces its source in only 0.28 of leaks. A fingerprint is a tool with physical bounds, and the rules for using it must be written in terms of batches, positions and the strength of rewording.

Scope of the numbers: But machine screening is a recall-first screen: on the book's synthetic labelled set only 0.380 of reworded risky clauses are caught, and the costlier a miss, the more a lawyer must read in full. Parameters are illustrative and all calculations are generated by the book's programs.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 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
Show the full program
# 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))

Single source and citation

Fan, Y. (2026). Human-AI Joint Asset and IP Protection, Automated Contract-Loophole Auditing and Full-Spectrum Anti-Erosion Architecture (Version 1.0.0), Chapter 4: Hidden Fingerprints: Floors on Code Length and the Decay of Rewording. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 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}
}

Other chapter answers in this book