Unverified50% confidenceBenchmarkExact time
Fine-tuned v3 achieved a safety recall of 13/21 (62%) at the 0.8 alert threshold, versus 0/21 for zero-shot; false alarms on ordinary chat dropped from 2.9% to 1.4%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/18/2026
First Seen
Valid until: 1/16/2027
Sources
Related Entities
Related Claims
UnverifiedIn the 5-agent conversational cluster architecture (unstructured Markdown handoffs), the spec drift failure rate is 48.4%, the silent tool error propagation rate is 79.0%, the premature termination rate is 34.2%, and the end-to-end success rate is 18.5%.69% similarUnverifiedAfter enabling reasoning mode, Qwen3.8-27B's bug detection rate dropped to 6/20, and 12 of 40 answers could not be parsed as JSON because they exhausted the 12,000-token output limit; Ornith 1.5's detection rate remained 8/20, with 13/40 answers unparseable, but it took about half the time of Qwen.69% similarUnverifiedIn the final round of experiments, the composite evaluation score improved by 0.0903, and zero-score conversations decreased by 40%.67% similarUnverifiedPrompt Guard 2 scored 0.999 on human-written injection instructions, but scored between 0.004 and 0.140 on real AgentDojo attacks, all below the default threshold of 0.5.66% similarUnverifiedIn the experiments, the task resolution rate was 17% for zero-shot generation, 27% for generic self-correction, 23% for error-only feedback, and 90% for diagnostic counterexample feedback (A-CEGIS).65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1062481API
curl https://kongchang.com/api/v1/knowledge/claims/1062481MCP
get_claim(id=1062481)