Unverified50% confidenceFactExact time
The 0-1 score obtained by normalizing Shieldstral's output top-20 token logprobs as exp(zYes)/(exp(zYes)+exp(zNo)) is a model score, not a calibrated true harm probability.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedApplying a logit bias of -2 reduces the probability of the target tokens to approximately 1/7 of the original (e^-2 ≈ 0.135).63% similarUnverifiedThe out-of-the-box graph tool returned zero callers for 17 out of 26 functions, despite every one of them having actual callers, and every empty result was labeled as precise with a risk level of LOW.63% similarUnverifiedThe log-score (i.e., cross-entropy) penalizes extreme errors more aggressively, and tends toward negative infinity when a model predicts the wrong answer with a probability close to 1.0.62% similarUnverified当z>4时，在零假设下观察到如此极端值的概率约为p≈3.2×10⁻⁵，即每十万次检测中约3次误报61% similarUnverifiedShannon pointed out that if an event has n equally probable possible outcomes, the minimum amount of information needed to determine the specific outcome is log₂(n) bits.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1034556API
curl https://kongchang.com/api/v1/knowledge/claims/1034556MCP
get_claim(id=1034556)