What does Human-AI Mutual Feeding and the Falsification Ratchet conclude, in numbers, about “Classical Sources: Teaching and Learning in the Xueji, Blue from Indigo in the Xunzi, Bacon's Idols and Popper's Conjectures”?

Answer: The closed loop is worse. If each round's consensus becomes the next round's prior, the shared bias accumulates round by round. With no outside anchor, the bias is 0.500 after 10 rounds, 2.500 after 50 and 10.000 after 200, with no upper limit. If a share of 0.05 of each round is anchored on outside evidence, the bias settles at 0.950; with an anchored share of 0.20, at 0.200; with 0.50, at 0.050, which is the single-round bias itself.

Scope of the numbers: Parameters are illustrative and all calculations are generated by the book's programs. Chapter 2 treats the jagged frontier: with illustrative parameters, clicking confirm outside the frontier is right 0.604 of the time against 0.800 for answering alone, while falsifying first and then comparing leads on both sides.

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: 0.05, 0.050, 0.20, 0.200, 0.50, 0.500

MSE of the average of k agents (shared bias 0.05, sd 1)
k MSE bias share of MSE
11.0025 0.002
30.3358 0.007
100.1025 0.024
300.0358 0.070
1000.0125 0.200
1,0000.0035 0.714
Closed-loop bias after t rounds
anchor a t10 t=50 t=200 limit
0.000.500 2.500 10.000 none
0.050.381 0.877 0.950 0.950
0.200.179 0.200 0.200 0.200
0.500.050 0.050 0.050 0.050
Show the full program
# Chapter 7. p72: Bacon's idols of the tribe as a shared bias.
# k agents estimate a quantity; each has private noise sd 1 and all
# share the same bias b0 = 0.05 (one base model, one training corpus).
# Averaging gives MSE = b0^2 + 1/k: more agents never remove b0.
# In a closed loop the consensus becomes the next round's prior, so the
# shared lean accumulates: b_{t+1} = (1 - a)(b_t + b0), where a is the
# share of each round anchored on outside evidence. With a = 0 the bias
# grows without limit; with a > 0 it settles at (1 - a) b0 / a.
# Illustrative parameters.
B0 = 0.05

if __name__ == "__main__":
    print("MSE of the average of k agents (shared bias 0.05, sd 1)")
    print("      k     MSE   bias share of MSE")
    for k in (1, 3, 10, 30, 100, 1000):
        mse = B0 * B0 + 1.0 / k
        print("  {:>5,}  {:.4f}  {:>17.3f}".format(k, mse, B0 * B0 / mse))
    print("Closed-loop bias after t rounds")
    print("  anchor a    t=10    t=50   t=200   limit")
    for a in (0.0, 0.05, 0.20, 0.50):
        b, row = 0.0, {}
        for t in range(1, 201):
            b = (1.0 - a) * (b + B0)
            row[t] = b
        lim = "none" if a == 0.0 else "{:.3f}".format((1 - a) * B0 / a)
        print("  {:>8.2f}  {:>6.3f}  {:>6.3f}  {:>6.3f}  {:>6}".format(
            a, row[10], row[50], row[200], lim))

Single source and citation

Fan, Y. (2026). Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding (Version 1.0.0), Chapter 7: Classical Sources: Teaching and Learning in the Xueji, Blue from Indigo in the Xunzi, Bacon's Idols and Popper's Conjectures. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-374 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026374,
  author    = {Fan, Yafu},
  title     = {Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book