What does Human-AI Mutual Feeding and the Falsification Ratchet conclude, in numbers, about “The Falsification Gate: Sequential Boundaries, Sample Sizes and the Leak Rate”?
Answer: Chapter 4 builds the falsification gate: a sequential test needs 98 to 118 cases on average where a fixed design needs 211 at the same error rates, and the leak rate is computed by candidate source.
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.
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: 118, 211, 98
| SPRT boundaries for alpha = 0.05, beta = 0.10: | |||||||
| ln A = 2.8904 (A = 18.0), ln B = -2.2513 (B = 0.1053) | |||||||
| p0 | p1 | alpha | beta | E0[N] | E1[N] | fixed n | saving |
| 0.5 | 0.6 | 0.05 | 0.10 | 98 | 118 | 211 | 0.441 |
| 0.5 | 0.6 | 0.01 | 0.05 | 143 | 207 | 388 | 0.465 |
| 0.5 | 0.7 | 0.05 | 0.10 | 23 | 29 | 50 | 0.422 |
| 0.5 | 0.7 | 0.01 | 0.05 | 33 | 51 | 92 | 0.448 |
| 0.3 | 0.4 | 0.05 | 0.10 | 92 | 105 | 191 | 0.449 |
| 0.3 | 0.4 | 0.01 | 0.05 | 135 | 185 | 351 | 0.473 |
| 0.7 | 0.8 | 0.05 | 0.10 | 71 | 92 | 161 | 0.426 |
| 0.7 | 0.8 | 0.01 | 0.05 | 103 | 162 | 298 | 0.455 |
Show the full program
# Chapter 4. p41: Wald's SPRT for a candidate rule, exact formulas.
# A rule claims that a check passes with rate p1 instead of the base
# rate p0. Each back-filled case is a Bernoulli trial. Wald (1945):
# upper boundary ln A = ln((1 - beta) / alpha) -> accept the rule
# lower boundary ln B = ln(beta / (1 - alpha)) -> reject the rule
# Average sample numbers, ignoring overshoot:
# E0[N] = ((1 - alpha) ln B + alpha ln A) / E0[z]
# E1[N] = (beta ln B + (1 - beta) ln A) / E1[z]
# with z = x ln(p1/p0) + (1 - x) ln((1 - p1)/(1 - p0)). The fixed
# sample size with the same errors is
# n = ((z_a sqrt(p0 q0) + z_b sqrt(p1 q1)) / (p1 - p0))^2.
import math
from statistics import NormalDist
Z = NormalDist()
def wald(p0, p1, a, b):
la = math.log((1.0 - b) / a)
lb = math.log(b / (1.0 - a))
up = math.log(p1 / p0)
dn = math.log((1.0 - p1) / (1.0 - p0))
e0 = p0 * up + (1.0 - p0) * dn
e1 = p1 * up + (1.0 - p1) * dn
n0 = ((1.0 - a) * lb + a * la) / e0
n1 = (b * lb + (1.0 - b) * la) / e1
return la, lb, n0, n1
def fixed_n(p0, p1, a, b):
za, zb = Z.inv_cdf(1.0 - a), Z.inv_cdf(1.0 - b)
num = za * math.sqrt(p0 * (1 - p0)) + zb * math.sqrt(p1 * (1 - p1))
return math.ceil((num / (p1 - p0)) ** 2)
if __name__ == "__main__":
print("SPRT boundaries for alpha = 0.05, beta = 0.10:")
la, lb, _, _ = wald(0.5, 0.6, 0.05, 0.10)
print(" ln A = {:.4f} (A = {:.1f}), ln B = {:.4f} (B = {:.4f})".format(
la, math.exp(la), lb, math.exp(lb)))
print(" p0 p1 alpha beta E0[N] E1[N] fixed n saving")
for p0, p1 in ((0.5, 0.6), (0.5, 0.7), (0.3, 0.4), (0.7, 0.8)):
for a, b in ((0.05, 0.10), (0.01, 0.05)):
_, _, n0, n1 = wald(p0, p1, a, b)
nf = fixed_n(p0, p1, a, b)
sav = 1.0 - max(n0, n1) / nf
print(" {:.1f} {:.1f} {:>5.2f} {:>4.2f} {:>5.0f} {:>5.0f}"
" {:>7} {:>6.3f}".format(p0, p1, a, b, n0, n1, nf,
sav))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 4: The Falsification Gate: Sequential Boundaries, Sample Sizes and the Leak Rate. 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
- Chapter 1 · Variance Collapse under Self-Feeding: Why the Tails Go First
- Chapter 2 · The Jagged Frontier: Why Clicking Confirm Loses Accuracy, and Why Better Individuals Make a Narrower Crowd
- Chapter 3 · The Gain from Injected Outliers: How Worse-on-Average Candidates Raise the Best
- Chapter 5 · A Robust Ratchet: Random Walks with Misclassification, Rollback Audits and the Condition for Compounding
- Chapter 6 · Multi-Agent Falsification: Majority Error and the Effective Number of Votes under Correlation
- Chapter 7 · Classical Sources: Teaching and Learning in the Xueji, Blue from Indigo in the Xunzi, Bacon's Idols and Popper's Conjectures
- Chapter 8 · Field Tools and Checker Specifications: the Daily Ratchet Ledger, Red-Yellow-Green Criteria and a Pre-registered Refutation