What does Human-AI Mutual Feeding and the Falsification Ratchet conclude, in numbers, about “The Gain from Injected Outliers: How Worse-on-Average Candidates Raise the Best”?

Answer: Chapter 3 computes the gain from injected human outliers: a pool of 20 model candidates has an expected best of 0.934, two human hypotheses lift it to 1.186, and the gain holds only when the hypotheses are testable.

Scope of the numbers: Parameters are illustrative and all calculations are generated by the book's programs. Chapter 2 treats the jagged frontier: with illustrative parameters, clicking confirm outside the frontier is right 0.604 of the time against 0.800 for answering alone, while falsifying first and then comparing leads on both sides.

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: 0.934, 1.186, 20

Pool of 20 candidates; k of them injected by humans
kshareE[best]P(best > 1.5)
00.000.9340.027
10.051.0660.114
20.101.1860.194
40.201.3970.332
60.301.5750.447
100.501.8570.621
150.752.1130.763
201.002.3010.852
Single candidate means: model 0.00, human -0.50
Show the full program
# Chapter 3. p31: the expected best of a mixed candidate pool.
# A daily pool holds N = 20 candidate rules. Model candidates have true
# value ~ N(0, 0.5^2): competent and alike. Human field hypotheses
# ~ N(-0.5, 1.5^2): worse on average, far more spread. With k human
# candidates the best value M has CDF
#   F(x) = Phi((x + 0.5) / 1.5)^k * Phi(x / 0.5)^(N - k),
# and E[M] = int_0^inf (1 - F) dx - int_-inf^0 F dx (trapezoid rule on
# [-12, 12]). Also P(M > 1.5), a "breakthrough" no model candidate
# reaches often. Illustrative parameters.
import math
from statistics import NormalDist

N = 20
AI = NormalDist(0.0, 0.5)
HU = NormalDist(-0.5, 1.5)
STEP = 0.001


def cdf_max(x, k):
    return HU.cdf(x) ** k * AI.cdf(x) ** (N - k)


def expected_max(k):
    pos, neg = [], []
    steps = int(12.0 / STEP)
    for i in range(steps + 1):
        x = i * STEP
        wgt = 0.5 if i in (0, steps) else 1.0
        pos.append(wgt * (1.0 - cdf_max(x, k)))
        neg.append(wgt * cdf_max(-x, k))
    return (math.fsum(pos) - math.fsum(neg)) * STEP


if __name__ == "__main__":
    print("Pool of {} candidates; k of them injected by humans".format(N))
    print("   k  share   E[best]   P(best > 1.5)")
    rows = []
    for k in (0, 1, 2, 4, 6, 10, 15, 20):
        e = expected_max(k)
        p = 1.0 - cdf_max(1.5, k)
        rows.append((k, e))
        print("{:>4}  {:>5.2f}  {:>8.3f}  {:>14.3f}".format(k, k / N, e, p))
    print("Single candidate means: model {:.2f}, human {:.2f}".format(
        AI.mean, HU.mean))

Single source and citation

Fan, Y. (2026). Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding (Version 1.0.0), Chapter 3: The Gain from Injected Outliers: How Worse-on-Average Candidates Raise the Best. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-374 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026374,
  author    = {Fan, Yafu},
  title     = {Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book