What does Antifragile Reputation and Audience Sovereignty conclude, in numbers, about “Malicious Reviews Leave Clues, Not Verdicts: Text Overlap, Bunching in Time and Base Rates”?

Answer: Chapter 4 shows that clues of malice are only clues: with all three clues present and a base rate of 0.05, the posterior probability of malice is only 0.341.

Scope of the numbers: Its parameters are illustrative; all calculations are generated by its programs; it is general business guidance, not legal advice. Marked boxes separate published evidence, runnable Python calculations (22 programs), classical sources and the author's hypotheses.

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: 0.05, 0.341

A. P(malicious | clues) under independence
clues LR base 0.020.05 0.15
young3.0 0.058 0.136 0.346
overlap6.0 0.109 0.240 0.514
burst2.5 0.049 0.116 0.306
young+overlap18.0 0.269 0.486 0.761
young+burst7.5 0.133 0.283 0.570
overlap+burst15.0 0.234 0.441 0.726
young+overlap+burst45.0 0.479 0.703 0.888
B. all three clues, correlated (rho 0.6): LR 9.8
base 0.02: P(malicious) 0.167; genuine per 100 flags 83.3
base 0.05: P(malicious) 0.341; genuine per 100 flags 65.9
base 0.15: P(malicious) 0.634; genuine per 100 flags 36.6
C. LR needed for P(malicious) >0.95 at each base rate
base 0.02: LR >931
base 0.05: LR >361
base 0.15: LR >108
Show the full program
# Chapter 4. p43: why a clue score is a lead, never a verdict.
# Three clues on an extreme negative review, each with an illustrative
# likelihood ratio for "malicious" against "genuine": young account
# (LR 3.0), text overlap with another review above the p41 threshold
# (LR 6.0), inside a burst window of p42 (LR 2.5). Naive independence
# multiplies them; correlated clues are discounted with an exponent
# RHO < 1 on the combined log LR. Base rates (share of extreme
# negatives that are malicious) are illustrative.
import itertools
import math

CLUES = (("young", 3.0), ("overlap", 6.0), ("burst", 2.5))
BASES = (0.02, 0.05, 0.15)
RHO = 0.6


def posterior(base, lr):
    odds = base / (1.0 - base) * lr
    return odds / (1.0 + odds)


if __name__ == "__main__":
    print("A. P(malicious | clues) under independence")
    print("   clues                   LR      base 0.02  0.05    0.15")
    for r in range(1, 4):
        for combo in itertools.combinations(CLUES, r):
            lr = math.prod(c[1] for c in combo)
            name = "+".join(c[0] for c in combo)
            print("   {:<22}  {:<6.1f}  {:.3f}      {:.3f}   {:.3f}".format(
                name, lr, *[posterior(b, lr) for b in BASES]))
    full = math.prod(c[1] for c in CLUES)
    disc = math.exp(RHO * math.log(full))
    print("B. all three clues, correlated (rho {:.1f}): LR {:.1f}".format(
        RHO, disc))
    for b in BASES:
        print("   base {:.2f}: P(malicious) {:.3f}; genuine per 100"
              " flags {:.1f}".format(
                  b, posterior(b, disc), 100.0 * (1.0 - posterior(b, disc))))
    print("C. LR needed for P(malicious) >= 0.95 at each base rate")
    for b in BASES:
        need = 0.95 / 0.05 * (1.0 - b) / b
        print("   base {:.2f}: LR >= {:,.0f}".format(b, need))

Single source and citation

Fan, Y. (2026). Antifragile Reputation Judo, Evidence-Backed Smear Neutralisation and Platform-Independent Audience Sovereignty (Version 1.0.0), Chapter 4: Malicious Reviews Leave Clues, Not Verdicts: Text Overlap, Bunching in Time and Base Rates. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-403 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026403,
  author    = {Fan, Yafu},
  title     = {Antifragile Reputation Judo, Evidence-Backed Smear Neutralisation and Platform-Independent Audience Sovereignty},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book