SOURCE_AUDIT.md

# Source audit

- Paper: [arXiv 2601.23221](https://arxiv.org/abs/2601.23221).
- OpenReview: [niQUth28zn](https://openreview.net/forum?id=niQUth28zn).
- No official GitHub repository is linked in the paper source; this reproduction is clean-room.

The paper calls its majority-skill condition an interpretable **sufficient** condition. Anchored Claim 4 changes this to a necessity statement (“requires …; without this … fails”), which is stronger and false in full generality: a homogeneous adversarial crowd can converge to `1-Y`, whose absolute demographic-parity gap equals that of `Y`. The audit reports this qualification directly.

All numerical experiments use the paper's finite binary `X=empty` setup. Majority/Bayes quantities are exact binomial sums, and FairCrowd optimality is checked through an independently formulated primal linear program over randomized finite-domain classifiers.