it avoids extreme probabilities more than the market does. When the true probability is 90%, the market might say 87% while the LLM says 78%. This compression hurts the LLM's Brier score on questions with extreme true probabilities.