What does Drift Control for Multi-Window Writing conclude, in numbers, about “Method: Five Failure Modes, Exclusive Locks and the Four Methods of Collation”?
Answer: Chapter 1 is the method. When the windows' errors are correlated, adding windows does not average them away: at an intraclass correlation of 0.02, sixteen windows are worth only 48.35 independent checks. Four windows sharing one stock of boilerplate should expect 18.22 duplicate pairs in a nine-chapter book, against 4.55 with separate stocks; a "check, then create" lock ends in double ownership in 2160 of the 2520 interleavings of four windows, and an exclusive create brings that to zero; at the error rates a published study found for GPT-4 citations, a chapter with 12 citations has a probability of only 0.003432 that all are real and correct.
Scope of the numbers: Every error rate without a source is an assumed parameter or a measurement on synthetic data; the book is not legal advice and recommends no writing tool or model. Marked boxes separate published evidence, runnable calculations, sources and the author's hypotheses, each stating what would refute it.
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: 0.02, 18.22, 4.55, 48.35
| A. effective independent checks N_eff = N / (1 + (N - 1) rho) | |
| windows checks N rho | 0.00 rho=0.02 rho=0.05 |
| 1 | 90 90.0 32.37 16.51 |
| 2 | 180 180.0 39.30 18.09 |
| 4 | 360 360.0 44.01 19.00 |
| 8 | 720 720.0 46.81 19.49 |
| 16 | 1440 1440.0 48.35 19.74 |
| ceiling 1/rho: rho | 0.02 -> 50, rho=0.05 -> 20 |
| B. boilerplate collisions: T | 2700 sentences, K = 20 stock lines, q = 0.01 |
| C(T,2) | 3,643,650 |
| shared stock, expected duplicate pairs | 18.22 |
| separate stock per window (m | 4): 4.55; ratio 4.00 |
| 3 chapters, T | 900: expected pairs 2.02 |
| 9 chapters, T | 2700: expected pairs 18.22 |
| 27 chapters, T | 8100: expected pairs 164.00 |
| C. simulation, 400 books, seed 152: shared 18.12, separate 4.67 | |
| D. P(no duplicate) > | 0.95 needs E[pairs] <= 0.0513 |
| so q < | 0.000531: about one boilerplate sentence per 1,885 sentences |
Show the full program
# Chapter 1. p10: why more windows do not average errors out, and why shared templates breed duplicates.
# Part A. N checkable spots (terms, numbers, citations) are written by m windows. If the windows share
# one prompt template, their mistakes are correlated with intraclass correlation rho. The variance of
# the book-level error rate is p(1 - p) / N * (1 + (N - 1) rho), so the effective number of independent
# checks is N_eff = N / (1 + (N - 1) rho), which can never exceed 1 / rho however many windows write.
# Part B. Each sentence is, with probability q, drawn from a stock of K boilerplate sentences.
# With one shared stock, any two of the T sentences of the book can collide: E[pairs] = C(T, 2) q^2 / K.
# With a separate stock per window only pairs inside one window collide: m C(T / m, 2) q^2 / K.
# Part C. A seeded simulation checks Part B; Part D finds the largest q that keeps P(no duplicate)
# at 0.95 or more, using P(no duplicate) = exp(-E[pairs]).
import math
import random
SPOTS_PER_WINDOW = 90
RHOS = (0.0, 0.02, 0.05)
def n_eff(n, rho):
return n / (1 + (n - 1) * rho)
def pairs_shared(t, q, k):
return math.comb(t, 2) * q * q / k
def pairs_separate(t, q, k, m):
return m * math.comb(t // m, 2) * q * q / k
def simulate(t, q, k, m, shared, trials, rng):
dups = []
per = t // m
for _ in range(trials):
counts = {}
for w in range(m):
for _ in range(per):
if rng.random() < q:
key = rng.randrange(k) if shared else (w, rng.randrange(k))
counts[key] = counts.get(key, 0) + 1
dups.append(sum(c * (c - 1) // 2 for c in counts.values()))
return math.fsum(dups) / trials
if __name__ == "__main__":
print("A. effective independent checks N_eff = N / (1 + (N - 1) rho)")
print(" windows checks N rho=0.00 rho=0.02 rho=0.05")
for m in (1, 2, 4, 8, 16):
n = m * SPOTS_PER_WINDOW
print(" {:7d} {:8d} {:9.1f} {:9.2f} {:9.2f}".format(
m, n, *(n_eff(n, r) for r in RHOS)))
print(" ceiling 1/rho: rho=0.02 -> {:.0f}, rho=0.05 -> {:.0f}".format(1 / 0.02, 1 / 0.05))
t, k, q, m = 2700, 20, 0.01, 4
print("B. boilerplate collisions: T = {} sentences, K = {} stock lines, q = {}".format(t, k, q))
es, ep = pairs_shared(t, q, k), pairs_separate(t, q, k, m)
print(" C(T,2) = {:,}".format(math.comb(t, 2)))
print(" shared stock, expected duplicate pairs {:.2f}".format(es))
print(" separate stock per window (m = {}): {:.2f}; ratio {:.2f}".format(m, ep, es / ep))
for chapters in (3, 9, 27):
tt = chapters * 300
print(" {:2d} chapters, T = {:5d}: expected pairs {:8.2f}".format(
chapters, tt, pairs_shared(tt, q, k)))
rng = random.Random(152)
ss = simulate(t, q, k, m, True, 400, rng)
sp = simulate(t, q, k, m, False, 400, rng)
print("C. simulation, 400 books, seed 152: shared {:.2f}, separate {:.2f}".format(ss, sp))
lam = -math.log(0.95)
qmax = math.sqrt(lam * k / math.comb(t, 2))
print("D. P(no duplicate) >= 0.95 needs E[pairs] <= {:.4f}".format(lam))
print(" so q <= {:.6f}: about one boilerplate sentence per {:,.0f} sentences".format(qmax, 1 / qmax))Single source and citation
Fan, Y. (2026). Drift Control for Parallel Multi-Window Book Writing, Cross-Volume Terminology Graph Audits and Zero-Hallucination Citation Verifiers (Version 0.1.0), Chapter 1: Method: Five Failure Modes, Exclusive Locks and the Four Methods of Collation. 5D Library. DOI pending
DOI pending · monograph page goes live after publication · 5DL-B-152 · v0.1.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026152,
author = {Fan, Yafu},
title = {Drift Control for Parallel Multi-Window Book Writing, Cross-Volume Terminology Graph Audits and Zero-Hallucination Citation Verifiers},
year = {2026},
publisher = {5D Library},
edition = {0.1.0},
note = {DOI pending}
}Other chapter answers in this book
- Chapter 2 · Concurrency Control: Exclusive-Create Locks, Leases, Version Numbers and Wave Scheduling
- Chapter 3 · The Terminology Graph: Lock Tables, Synonym Drift and Cross-Volume Consistency Audits
- Chapter 4 · Number Parity and Duplicate Sentences: Bilingual Agreement, Program Reconciliation and Near-Duplicate Detection
- Chapter 5 · A Zero-Hallucination Citation Verifier: Fields, Identifiers, Ledger Matching and Item-by-Item Verdicts
- Chapter 6 · Style Drift: Gates for Sentence Length, Lexical Diversity, Stock Phrases and Paragraph Openers
- Chapter 8 · Practice: Four Windows Write a Nine-Chapter Book, Five Classes of Injected Error and the Detection Ledger of Four Gates