Condorcet's Jury Theorem with Correlated Errors: When Multi-Model Voting Gets More Accurate and When Shared Origins Make Everyone Wrong Together: chapter answers
Many owners let several artificial-intelligence models vote, or ask one model several times and take the majority, believing that a few more votes will make a verdict safe. The practice rests on Condorcet's jury theorem of 1785: if each vote is right with probability above one half and the votes are independent, the more votes there are, the more likely the majority is right. This book overturns the popular version of that theorem in multi-model voting: when models share origins and their errors are correlated, majority accuracy has a ceiling below 1. At single-vote accuracy 0.70, five indepe…
DOI pending · monograph page goes live after publication · 5DL-B-309 · 中文版
- Chapter 1 · Condorcet's Binomial Majority: Why Independent Votes Grow More Accurate
The five columns of panel A correspond to single-vote accuracies of 0.45, 0.55, 0.60, 0.70 and 0.80. Take the column for 0.70. One vote gives 0.7000, three give 0.7840, five give 0.8369, fifteen give 0.9500 and fifty-on… - Chapter 2 · Correlated Errors and the Ceiling: the Beta-Binomial Model, Opinion Leaders and Shared Origins
Whenever errors share a source, majority accuracy has a ceiling below 1. In the Beta-binomial model the ceiling is the share of questions on the right side of one half; in the opinion-leader model it depends on whether… - Chapter 3 · Measuring Correlation: Pairwise Correlation, the Q Statistic, Effective Votes and Colliding Wrong Answers
Correlation can be measured, and it should be. The phi coefficient and the moment estimate correspond directly to $\rho$ in the model; the Q statistic is more sensitive but runs on a different scale; the both-wrong shar… - Chapter 4 · Comparing Ways to Decorrelate: Resampling, Different Prompts, Different Models and Standard-Library Checks
Chapter 4 compares decorrelation strategies: resampling one model has a ceiling of only 0.7366, five different models a ceiling of 0.8257, and four models plus a standard-library check that decides wherever it applies r… - Chapter 5 · Weighted Majorities: Log-Odds Weights and the Double Counting of Correlated Clusters
Under independence, log-odds weighting is the optimal rule, and when one vote's weight exceeds the sum of all others it should decide alone, which gives Chapter 4's deciding check its theoretical footing. Under correlat… - Chapter 6 · How Many Models and How Many Labelled Questions: the Vote Ledger and the Labelling Ledger below the Ceiling
Under correlation the vote ledger fills with "never": the higher the target, the less correlation it tolerates, and beyond the threshold no number of extra calls helps. The labelling ledger reminds the owner that all th… - Chapter 7 · Classical Sources East and West: "When Three Read, Follow the Two" in the Great Plan and Condorcet's Essai of 1785
Chapter 7 examines "when three read, follow the two" in the Great Plan chapter of the Book of Documents and Condorcet's Essai, analysing the voting structure only and making no reading of any kind: under independence a… - Chapter 8 · Closing the Loop on Decorrelated Multi-Model Verdicts: a Procedure, the Correlated Voting Accuracy Calculator and a Pre-Registered Plan of Refutation
Chapter 8 gives the empirical loop for decorrelated multi-model verdicts, the distilled Correlated Voting Accuracy Calculator and a pre-registered plan of refutation. Marked boxes separate published evidence, runnable P…
Illustrative model calculation, not a live quote