According to Anthropic (@AnthropicAI), the scaling of 'elicitation probabilities' allows for accurate risk forecasts in model deployments, significantly reducing the need for evaluation queries.