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.

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: 0.02, 18.22, 4.55, 48.35

A. effective independent checks N_eff = N / (1 + (N - 1) rho)
windows checks N rho0.00 rho=0.02 rho=0.05
190 90.0 32.37 16.51
2180 180.0 39.30 18.09
4360 360.0 44.01 19.00
8720 720.0 46.81 19.49
161440 1440.0 48.35 19.74
ceiling 1/rho: rho0.02 -> 50, rho=0.05 -> 20
B. boilerplate collisions: T2700 sentences, K = 20 stock lines, q = 0.01
C(T,2)3,643,650
shared stock, expected duplicate pairs18.22
separate stock per window (m4): 4.55; ratio 4.00
3 chapters, T900: expected pairs 2.02
9 chapters, T2700: expected pairs 18.22
27 chapters, T8100: 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