Unverified50% confidenceFactExact time
The second stage used 72 test cases with ground-truth answers to measure each framework's ability to detect known risks.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/11/2026
First Seen
Valid until: 2/9/2027
Sources
Related Entities
Related Claims
UnverifiedIn the testing phase, the team chose the second route: humans first determine test scope and risk level, then the model scans the code to supplement steps; the process includes collecting affected interfaces from traffic monitoring and code changes, having the model read the code to answer the scale of changes and branches, and using decision tables to split first and then combine to generate streamlined test case combinations.74% similarUnverifiedThe study validated findings through capability thresholds and transplanted the audit methodology to a second self-discovered rule, OPEN, demonstrating the generalizability of the approach.74% similarUnverifiedThe researchers ran the MARCH framework as-is on two code evaluation benchmarks, covering 80 condition-by-unit measurements.74% similarUnverifiedAnswer evaluation is an independent second stage, requiring scoring criteria to test the factual grounding, citation targeting, question answering, refusal behavior, and access control compliance of generated answers.73% similarUnverifiedA second engineering implementation recommendation: when evaluating fine-tuned checkpoints, not only test the target capability B₁ but also test against a decoupled matrix of adversarial behaviors, making 'whether unrelated capability drift appears after fine-tuning' a routine metric.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1158580API
curl https://kongchang.com/api/v1/knowledge/claims/1158580MCP
get_claim(id=1158580)