What does Reasoning Trajectories and Outcome Luck conclude, in numbers, about “The Cost of Forgery: Probing Depth, Memory Load and Bounds on Detection”?

Answer: Chapter 5 models forgery cost: under illustrative parameters, probing to level six flags a performer with probability 0.784 and a genuine reasoner with 0.241, while the claim fails where scripts are ample and reasoning trees narrow, so a flag can only trigger review.

Scope of the numbers: After-the-fact explanations are often folk theories: in a synthetic experiment, retrospective self-reports correlate only -0.013 with the weights used but 0.826 with an observer's guess; concurrent thinking aloud keeps order and leaps. Parameters are illustrative and all calculations come from the book's 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. printed values that match the answer: 0.241, 0.784

S = 40, b = 3, q = 0.03, q_h = 0.003, notice = 0.7
depthcoverloadP(flag|perf)P(flag|genuine)ratio
11.00040.0000.0040.0
21.00080.0000.0250.0
31.000120.0000.0610.0
40.494160.2180.1112.0
50.165200.5440.1723.2
60.055240.7840.2413.3
70.018280.9130.3162.9
80.006320.9700.3932.5
Show the full program
# Chapter 5. p51: forgery cost against probing depth (illustrative).
# Two "why" probes are asked at each depth k = 1..D. A performer has a
# prepared script of S = 40 nodes; the reasoning tree branches b = 3
# ways, so the script covers depth k with prob c_k = min(1, S / b^k).
# Off script, the performer must improvise an answer consistent with
# every commitment made so far (two per answer); each check fails with
# prob q = 0.03. A genuine reasoner regenerates answers from a working
# model and slips with prob q_h = 0.003 per commitment. An assessor
# notices an inconsistency with prob 0.7. P(flag) = prob of at least
# one noticed inconsistency by depth D. Memory load = commitments.
S, B, Q, QH, SENS = 40, 3, 0.03, 0.003, 0.7
PER_DEPTH, PER_ANSWER = 2, 2


def flag_curve(s, b, q, qh, sens, depth):
    clean_p, clean_g, made = 1.0, 1.0, 0
    rows = []
    for k in range(1, depth + 1):
        cover = min(1.0, s / b ** k)
        for _ in range(PER_DEPTH):
            slip_p = 1 - (1 - q) ** made
            slip_g = 1 - (1 - qh) ** made
            clean_p *= 1 - (1 - cover) * slip_p * sens
            clean_g *= 1 - slip_g * sens
            made += PER_ANSWER
        rows.append((k, cover, made, 1 - clean_p, 1 - clean_g))
    return rows


def main():
    print("S = {}, b = {}, q = {}, q_h = {}, notice = {}".format(
        S, B, Q, QH, SENS))
    print("depth  cover  load  P(flag|perf)  P(flag|genuine)  ratio")
    for k, c, m, fp, fg in flag_curve(S, B, Q, QH, SENS, 8):
        ratio = fp / fg if fg > 0 else float("inf")
        print("{:>5}{:>7.3f}{:>6}{:>14.3f}{:>17.3f}{:>7.1f}".format(
            k, c, m, fp, fg, ratio))


main()

Single source and citation

Fan, Y. (2026). Reasoning-Trajectory Profiling, Process-Tracing Games and Separating Cognitive Structure from Stated Intentions and Outcome Luck (Version 1.0.0), Chapter 5: The Cost of Forgery: Probing Depth, Memory Load and Bounds on Detection. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-385 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026385,
  author    = {Fan, Yafu},
  title     = {Reasoning-Trajectory Profiling, Process-Tracing Games and Separating Cognitive Structure from Stated Intentions and Outcome Luck},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book