What does Bicameral AI Architecture conclude, in numbers, about “Two-Way Correction: Gain, Noise and the Limits of Self-Correction”?

Answer: Chapter 6 gives the stable interval of the correction gain and shows how it narrows when receipts arrive a round late; self-reflection without an outside signal pulls accuracy towards 0.6250, receipt-driven revision towards 0.9600.

Scope of the numbers: Its parameters are illustrative; all calculations are generated by its programs; "bicameral" is an engineering metaphor and no model is claimed to have awareness; the book is not personalised business or engineering advice. Marked boxes separate published evidence, runnable Python calculations (23 programs), classical sources and the author's hypotheses.

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: 0.6250, 0.9600

A. start accuracy 0.75
   round  self-review  receipts(sens 0.90)  receipts(sens 0.50)
   0      0.7500       0.7500               0.7500
   1      0.7300       0.8287               0.7888
   2      0.7132       0.8780               0.8192
   3      0.6991       0.9087               0.8430
   4      0.6872       0.9280               0.8618
   5      0.6773       0.9400               0.8765
   6      0.6689       0.9475               0.8881
B. fixed points: self-review 0.6250
   receipts 0.9600 (sens 0.90)
   receipts 0.9302 (sens 0.50)
C. after 6 receipt rounds: wrong 0.0525, of which unflagged by
   the last test 0.0053
Show the full program
# Chapter 6. p63: self-review without a signal versus test receipts.
# Two-state chain (answer right / wrong), start accuracy A0.
# Intrinsic self-review: no outside signal; each round a wrong answer
# becomes right with prob FIX and a right one becomes wrong with prob
# BREAK. Receipt-driven review: a test flags a wrong answer with
# sensitivity SENS and a right one with false-alarm rate FA; only a
# flagged answer is revised (wrong -> right with prob FIX_F, right ->
# wrong with prob BREAK_F). Exact recursion. All rates illustrative;
# they are not measurements from any published experiment.
A0 = 0.75
FIX, BREAK = 0.10, 0.06
SENS, FA, FIX_F, BREAK_F = 0.90, 0.05, 0.40, 0.30


def intrinsic(a):
    return a * (1.0 - BREAK) + (1.0 - a) * FIX


def receipt(a, sens=SENS):
    stay_right = a * (1.0 - FA * BREAK_F)
    fixed = (1.0 - a) * sens * FIX_F
    return stay_right + fixed


if __name__ == "__main__":
    print("A. start accuracy {:.2f}".format(A0))
    print("   round  self-review  receipts(sens 0.90)  receipts(sens 0.50)")
    a1 = a2 = a3 = A0
    for t in range(0, 7):
        print("   {}      {:.4f}       {:.4f}               {:.4f}".format(
            t, a1, a2, a3))
        a1, a2, a3 = intrinsic(a1), receipt(a2), receipt(a3, 0.50)
    print("B. fixed points: self-review {:.4f}".format(FIX / (FIX + BREAK)))
    r_fix = SENS * FIX_F / (SENS * FIX_F + FA * BREAK_F)
    print("   receipts {:.4f} (sens 0.90)".format(r_fix))
    r_fix5 = 0.50 * FIX_F / (0.50 * FIX_F + FA * BREAK_F)
    print("   receipts {:.4f} (sens 0.50)".format(r_fix5))
    a = A0
    for _ in range(6):
        a = receipt(a)
    print("C. after 6 receipt rounds: wrong {:.4f}, of which unflagged by"
          .format(1.0 - a))
    print("   the last test {:.4f}".format((1.0 - a) * (1.0 - SENS)))

Single source and citation

Fan, Y. (2026). Bicameral AI Architecture: Fusing Long-Context Holographic Strategy Models with Local Deterministic Execution Agents (Version 1.0.0), Chapter 6: Two-Way Correction: Gain, Noise and the Limits of Self-Correction. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-410 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026410,
  author    = {Fan, Yafu},
  title     = {Bicameral AI Architecture: Fusing Long-Context Holographic Strategy Models with Local Deterministic Execution Agents},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book