Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding: chapter answers
Three claims about writing with machines are widely repeated: a model that feeds on its own output can improve without limit; with a good model the human only has to click "confirm"; human intuition is unreliable, so everything should go to the data. This book refutes each claim or draws its boundary. Feeding a model its own output multiplies the expected variance by (n-1)/n per generation, and the tails vanish before the variance does; mixing in a real source gives the variance a steady state, and with 100 samples per generation the real share must be at least about 0.0909. After checking th…
DOI pending · monograph page goes live after publication · 5DL-B-374 · 中文版
- Chapter 1 · Variance Collapse under Self-Feeding: Why the Tails Go First
Self-feeding multiplies the expected variance by (n-1)/n per generation, typical paths shrink faster still, and the median half-life is about half the expected one. The tails leave before the variance does: a fifth less… - Chapter 2 · The Jagged Frontier: Why Clicking Confirm Loses Accuracy, and Why Better Individuals Make a Narrower Crowd
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… - Chapter 3 · The Gain from Injected Outliers: How Worse-on-Average Candidates Raise the Best
Chapter 3 computes the gain from injected human outliers: a pool of 20 model candidates has an expected best of 0.934, two human hypotheses lift it to 1.186, and the gain holds only when the hypotheses are testable. - Chapter 4 · The Falsification Gate: Sequential Boundaries, Sample Sizes and the Leak Rate
Chapter 4 builds the falsification gate: a sequential test needs 98 to 118 cases on average where a fixed design needs 211 at the same error rates, and the leak rate is computed by candidate source. - Chapter 5 · A Robust Ratchet: Random Walks with Misclassification, Rollback Audits and the Condition for Compounding
Chapter 5 treats the ratchet as a random walk with misclassification, adds rollback audits and states the condition for compounding: with illustrative parameters the feedback strength must exceed 0.0556. - Chapter 6 · Multi-Agent Falsification: Majority Error and the Effective Number of Votes under Correlation
Chapter 6 computes multi-agent falsification: correlated same-source agents add votes without adding protection, and the independent prediction of 0.0579 for five votes holds only under independence. - Chapter 7 · Classical Sources: Teaching and Learning in the Xueji, Blue from Indigo in the Xunzi, Bacon's Idols and Popper's Conjectures
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… - Chapter 8 · Field Tools and Checker Specifications: the Daily Ratchet Ledger, Red-Yellow-Green Criteria and a Pre-registered Refutation
Chapter 8 gives the field tools and checker specifications: a daily ratchet ledger, red-yellow-green criteria, the distilled Ratchet Gate and Collapse Alarm Calculator and a pre-registered refutation plan. Marked boxes…
Illustrative model calculation, not a live quote