Anti-Collapse Engineering with Two Anchors of Truth: the Variance-Collapse Rate of Self-Generated Training Data and Immunity through Standard-Library Recomputation and Textual Collation: chapter answers
When a model is retrained generation after generation on content it produced itself, the tails of the distribution disappear and its width collapses, which Shumailov and colleagues called model collapse in Nature in 2024. This book tests the founding claim that the variance collapses as sigma^2 (1-1/N)n and that mixing a fixed share of real sources into each generation prevents collapse, and corrects it after checking the originals: the formula is only the expectation under a maximum-likelihood refit, while the paper's theorem uses unbiased estimators, under which the expectation stays put an…
DOI pending · monograph page goes live after publication · 5DL-B-265 · 中文版
- Chapter 1 · Variance Collapsing Generation by Generation: the Expectation Formula of a Gaussian Feedback Chain and Almost-Sure Extinction
Chapters 1 and 2 derive the Gaussian feedback chain and the discrete absorbing chain: with 100 samples per generation the expected variance is down to 0.3660 at generation 100, and the typical value collapses about twic… - Chapter 2 · The Tails Die First: Generational Extinction in a Discrete Law and the Absorbing States of Wright–Fisher
Program p21 shows that a single round of finite sampling deletes a large stretch of the tail: with 1000 draws per generation, the Zipf list of 1000 categories is expected to lose 666.3 of them, the tail loses 0.8254 of… - Chapter 3 · Replace or Accumulate: the Counter-Evidence of Gerstgrasser and the Bound of pi^2/6
Chapter 3 sets out the contrary evidence of Gerstgrasser and colleagues: when real data are accumulated rather than replaced, the error of the mean in the Gaussian twin is capped by pi^2/6, that is 1.644934. - Chapter 4 · The Threshold of Mixing in Real Sources: How Much Variance a Fixed Share Preserves, and How Much of the Tail It Misses
Chapter 4 computes the fixed point of real-source mixing: at the same real share of 0.01 per generation, raising N from 100 to 10000 lifts the expected variance from 0.6644 to 0.9950; tail coverage answers only to count… - Chapter 5 · Examining for Collapse: Why Inward-Looking Metrics Go Blind, and How Much Data an Outward-Looking Anchor Needs
Chapter 5 shows that inward-looking detection metrics grow more optimistic during collapse, the Good–Turing missing mass falling from 0.0677 to 0.0009 while the true loss rises. - Chapter 6 · The First Anchor: Standard-Library Recomputation, Deterministic Fingerprints and the Correlated Failure of Two Anchors
Chapter 6 covers the first anchor: standard-library recomputation, the order independence of math.fsum and the correlated failure of two anchors, which together raise the probability of an error-free book of 300 numbers… - Chapter 7 · The Classics Examined: Duan Yucai on "the Difficulty of Settling Right and Wrong" and Erasmus's Return to the Original in 1516
Chapter 7 examines Duan Yucai's "difficulty of settling right and wrong" and Erasmus's New Testament of 1516. - Chapter 8 · The Empirical Loop: a Two-Anchor Procedure for Every Volume, the Feedback Collapse Rate Calculator and a Pre-Registered Test
Chapter 8 gives an eight-step two-anchor procedure for every volume, the distilled Feedback Collapse Rate Calculator and a pre-registered test. Marked boxes separate published evidence, runnable Python calculations (22…
Illustrative model calculation, not a live quote