What does Hard Gates and Soft-Prompt Decay conclude, in numbers, about “Decay over Turns: an Exponential Model of Soft-Prompt Compliance, Re-Sending Periods and Reading the Log”?

Answer: Chapter 3 gives an exponential decay model with a floor, with a half-life of 6.93 turns, under which a forty-turn session lets through 4.71 times the static estimate.

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: 4.71, 6.93

half-life of the excess compliance ln2/k6.93 turns
turn comply miss
00.9500 0.0500
50.8516 0.1484
100.7920 0.2080
200.7338 0.2662
400.7046 0.2954
800.7001 0.2999
expected escapes per session, decaying versus static model:
T10 decaying 0.134 static 0.050 ratio 2.68
T40 decaying 0.942 static 0.200 ratio 4.71
T100 decaying 2.737 static 0.500 ratio 5.47

Key formula

\[c(t) = c_\infty + (c_0 - c_\infty)\,e^{-kt}.\]
Show the full program
# Chapter 3. p31: a soft prompt whose pull fades with the turns.
# Author's working model (a hypothesis, not a published law): the
# chance that the agent follows a standing instruction at turn t is
#   c(t) = c_inf + (c0 - c_inf) x exp(-k t),
# starting at c0 and sinking to a floor c_inf. Illustrative values:
# c0 = 0.95, c_inf = 0.70, k = 0.10 per turn. Each turn offers a
# violation opportunity with probability v = 0.10, as in p11.
import math

C0, CINF, K, V = 0.95, 0.70, 0.10, 0.10


def comply(t):
    return CINF + (C0 - CINF) * math.exp(-K * t)


print("half-life of the excess compliance ln2/k = %.2f turns"
      % (math.log(2) / K))
print("%5s %10s %10s" % ("turn", "comply", "miss"))
for t in (0, 5, 10, 20, 40, 80):
    print("%5d %10.4f %10.4f" % (t, comply(t), 1 - comply(t)))
print("expected escapes per session, decaying versus static model:")
for T in (10, 40, 100):
    dec = math.fsum(V * (1 - comply(t)) for t in range(T))
    sta = V * (1 - C0) * T
    print("  T = %3d  decaying %.3f  static %.3f  ratio %.2f" % (
        T, dec, sta, dec / sta))

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 3: Decay over Turns: an Exponential Model of Soft-Prompt Compliance, Re-Sending Periods and Reading the Log. 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