Errors Extended to Maintain Consistency
為維持一致而延伸錯誤
Verification & Validation
Risk Description
When a system misreads a foundational figure at the beginning of an analytical report, and all subsequent sections build on it; due to unmitigated control gaps, the report is rigorously argued and structurally complete, so reviewers do not notice the flawed premise, triggering external stakeholder impacts and causing the error surfaces only at the execution stage.
Framework Mappings
NIST AI 600-1Confabulation
OWASP Top 10 for LLMLLM09
NIST AI RMFMEASURE 2.5
ISO/IEC 5338驗證與確效
MIT AI Risk RepositoryDomain 7
Risk Treatment & Implementation Guidance
Require base data and premises to be listed separately in analytical reports, verifying premises before the derivation in review; Perform stepwise fact verification on key reports, confirming base data against original sources; Use reasoning-chain checks to flag outputs that are formally coherent but premise-unverified