What does Hard Gates and Soft-Prompt Decay conclude, in numbers, about “The Product Account: Why Stacked Soft Layers Never Leave a Zero Residual”?

Answer: Soft reminders work in proportion; stacked, they leave a product, and with illustrative parameters four layers leave 0.004800, not zero. A hard gate sorts violations into an in-class part, which escapes only when the gate breaks and passes, and an out-of-class part, which it does not touch. Seeing zero yields only an upper bound; the guarantee of zero comes from structure. The product account, however, rests on a strong assumption: that the layers are independent. If every soft layer reads the same context and suffers the same lapse, will the holes open together, and where will the gain from stacking stop? That is the question of the next chapter.

Scope of the numbers: Soft reminders work in proportion; when layers are independent the residual violation rate is the product of their miss rates, and with illustrative parameters four layers still leave 0.0048, not zero. Its parameters are illustrative; all calculations are generated by its programs.

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: 0.004800

base violation share v = 0.10 per save
layer added    miss     residual  per 1e6 saves
prompt         0.30     0.030000        30000.0
self-check     0.50     0.015000        15000.0
peer           0.40     0.006000         6000.0
human          0.80     0.004800         4800.0
residual is exactly 3/625, never 0
layers of miss 0.30 needed to reach a target residual:
  target 1e-04   layers  6
  target 1e-06   layers 10
  target 1e-09   layers 16
a deterministic layer with miss 0 inside its class: residual 0

Key formula

\[R = v \times m_1 \times m_2 \times \cdots \times m_k.\]
Show the full program
# Chapter 1. p11: residual violation rate of stacked soft layers, exact.
# Model: an agent would commit one class of violation in a share v of
# its saves. Soft layers (a system prompt, a self-check reminder, a
# peer review, a sampled human look) each let a violation through with
# miss rate m_i. If the layers fail independently, the residual rate
# is v times the product of all m_i: it shrinks but never reaches
# zero while every m_i > 0. All rates are illustrative.
from fractions import Fraction as F
import math

V = F(1, 10)
LAYERS = [("prompt", F(3, 10)), ("self-check", F(1, 2)),
          ("peer", F(2, 5)), ("human", F(4, 5))]

print("base violation share v = %.2f per save" % float(V))
print("%-12s %6s %12s %14s" % ("layer added", "miss", "residual",
                                "per 1e6 saves"))
r = V
for name, m in LAYERS:
    r = r * m
    print("%-12s %6.2f %12.6f %14.1f" % (name, float(m), float(r),
                                          float(r) * 1e6))
print("residual is exactly %s, never 0" % r)
print("layers of miss 0.30 needed to reach a target residual:")
for target in (F(1, 10 ** 4), F(1, 10 ** 6), F(1, 10 ** 9)):
    k = math.ceil(math.log(float(target / V)) / math.log(0.3))
    print("  target %.0e   layers %2d" % (float(target), k))
print("a deterministic layer with miss 0 inside its class: residual 0")

Single source and citation

Fan, Y. (2026). Hard Gates across the Client Lifecycle: the Decay of Soft-Prompt Compliance over Turns and State Machines for Interception before Publication (Version 1.0.0), Chapter 1: The Product Account: Why Stacked Soft Layers Never Leave a Zero Residual. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-256 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026256,
  author    = {Fan, Yafu},
  title     = {Hard Gates across the Client Lifecycle: the Decay of Soft-Prompt Compliance over Turns and State Machines for Interception before Publication},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book