Unverified50% confidenceBenchmarkExact time
Finite-sample numerical experiments show that under both settings, the sample risk premium accounts for about 74% to 97% of the total regret.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the synthetic experiment, "always output RISK_LEVEL: 1" achieves an ordinary accuracy as high as 0.922, but an average reward of −0.000, showing that class-balanced rewards can prevent laziness.72% similarUnverifiedAn increase in relative risk may be negligible in absolute terms if the baseline incidence rate is already very low.71% similarUnverified校准的不确定性区间意味着：若模型声称 90% 置信区间为某范围，则在大量测试样本中约 90% 的真实值应落入该范围。71% similarUnverifiedWhen reporting success rates, the sample size must be included, because 19/20 and 950/1000 represent different levels of evidential strength.69% similarUnverifiedIn this case, the lightweight model handled about 65% of request volume but contributed only about 2.4% of the cost.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133661API
curl https://kongchang.com/api/v1/knowledge/claims/1133661MCP
get_claim(id=1133661)