What does The Limits of Guardrail Compression conclude, in numbers, about “Closing the Loop: One Compression-Recall Curve per Volume, a Knee Calculator and a Pre-Registered Protocol”?

Answer: Chapter 8 gives the loop of one compression-recall curve per volume, the distilled "Rule-Compression Knee Calculator" and a paired pre-registered protocol needing 75 to 103 drafts per test when the gap in accuracy is large. Marked boxes separate published evidence, runnable Python calculations (20 programs), classical sources and the author's hypotheses.

Scope of the numbers: Chapter 1 builds an illustrative source of eight kinds of draft and three verdicts and the counting bound: three rules keep 0.960 of the verdict information, two only 0.629. Chapter 4 shows that the critical exchange rate is the reciprocal of the squared maximal correlation, 0.6844 and 1.4610 on the illustrative source while information rises continuously as rules split.

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: 103, 75

differencediscordantn (power 0.80)n (power 0.90)
0.1270.2075103
0.1270.30113156
0.0500.10246339
0.0500.20493682
0.0200.057711067
0.0200.1015442138
block recall test (drop 0.05, discordant 0.08):
block share 0.100 ->196 block drafts,1960 drafts in all
block share 0.267 ->196 block drafts,735 drafts in all
block share 0.400 ->196 block drafts,490 drafts in all
Show the full program
# Chapter 8. p82: how many held-out drafts the pre-registered test needs.
# Two rule sets judge the same drafts: the knee set and the set with one
# rule fewer. Outcome per draft: does each set reach the reviewer's
# decision? Paired binary outcomes, McNemar-type sample size (normal
# approximation, Connor 1987 form): n = (z_a sqrt(d) + z_b sqrt(d - e^2))
# ^ 2 / e^2, d = share of discordant drafts, e = difference in accuracy.
# One-sided alpha 0.05, power 0.80 and 0.90. For recall, n counts drafts
# that should be blocked; divide by their share to get all drafts.
import math
from statistics import NormalDist

Z = NormalDist()


def paired_n(e, d, alpha=0.05, power=0.80):
    za, zb = Z.inv_cdf(1 - alpha), Z.inv_cdf(power)
    return math.ceil((za * math.sqrt(d) + zb * math.sqrt(d - e * e)) ** 2
                     / (e * e))


print("difference  discordant  n (power 0.80)  n (power 0.90)")
for e, d in [(0.127, 0.20), (0.127, 0.30), (0.05, 0.10), (0.05, 0.20),
             (0.02, 0.05), (0.02, 0.10)]:
    print("   %5.3f       %4.2f          %5d           %5d"
          % (e, d, paired_n(e, d), paired_n(e, d, power=0.90)))
print("block recall test (drop 0.05, discordant 0.08):")
n_block = paired_n(0.05, 0.08)
for share in [0.10, 0.267, 0.40]:
    print("  block share %5.3f -> %4d block drafts, %5d drafts in all"
          % (share, n_block, math.ceil(n_block / share)))

Single source and citation

Fan, Y. (2026). The Information Bottleneck and Rate-Distortion Limits: How Far a Bilingual Monograph Can Be Compressed into Guardrails before Discriminative Power Falls Away (Version 1.0.0), Chapter 8: Closing the Loop: One Compression-Recall Curve per Volume, a Knee Calculator and a Pre-Registered Protocol. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-250 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026250,
  author    = {Fan, Yafu},
  title     = {The Information Bottleneck and Rate-Distortion Limits: How Far a Bilingual Monograph Can Be Compressed into Guardrails before Discriminative Power Falls Away},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book