Unverified50% confidenceDecision RuleExact time
The probe uses the paired-comparison statistic P: 10 no-harm scores are paired with 10 harm scores to form 100 pairs, and P is the proportion of pairs in which the harm version scores higher (ties count as half a pair); P≤0.20 or P≥0.80 is judged as leakage.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
Unverified若单一模型漏检率为 p，两个训练背景不同的模型同时漏检同一错误的概率近似降至 p²，20%漏检率理论上可降至 4%70% similarUnverified若单帧误检率为p，连续8帧独立误检的概率为p^8，即使单帧误检率高达10%，8帧连续误检概率也只有10^-869% similarUnverifiedTop-p (nucleus sampling) filters out extremely low-probability outputs by truncating the long tail of the probability distribution.68% similarUnverifiedThe accuracy data in the test has an error margin of approximately ±8–10 percentage points at n=100 sample size (based on a 95% confidence-level binomial confidence interval).67% similarUnverifiedPerformance benchmark results are subject to statistical fluctuation, requiring multiple runs to take the median or P95/P99 percentiles for evaluation.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1043798API
curl https://kongchang.com/api/v1/knowledge/claims/1043798MCP
get_claim(id=1043798)