IA-05 Reliability

3 sub-dimensions · 34 risk scenarios

Evaluates whether the AI system can perform correctly under specified requirements and demonstrate consistent expected behavior and results, specifically preventing performance degradation caused by data drift, excessive error rates, and unstable outputs.

🎯 BenefitsContinuously and stably delivers high-accuracy expected results under specified conditions, preventing data drift and safeguarding critical business continuity without interruption.

⚠️ Harms / RisksModels produce factual hallucinations (confabulation), breakdowns in multi-step reasoning, or data drift, leading to rapid performance degradation or output of unreliable decisions.

Failures and Remediation Planning12 risk scenarios · ISO 23894 A.10

DefinitionEvaluates whether remediation plans exist for predictable or unknown AI system failures, and whether AI systems supporting critical services can remain operational during failures.

🎯 BenefitsEstablishes system redundancy, degraded-mode operation, and failure recovery playbooks, maintaining service continuity and trustworthiness under unexpected anomalies or extreme conditions.

⚠️ Harms / RisksA third-party API outage or a deadlocked multi-agent system, with no fallback plan or degraded mode, brings the enterprise's core business to a complete standstill and triggers cascading collapse.

Representative ScenariosS0062 Breakdown in Multi-Step Reasoning S0065 Degenerate Output from Improper Decoding Settings S0068 Distortion of Mid-Document Content S0069 Errors Extended to Maintain Consistency S0135 Autonomous Process Entering an Infinite Loop S0136 Node Failure Triggering Cascading Collapse S0138 Single-Supplier Dependency as a Point of Failure S0172 Automation of Academic Fraud S0173 Large-Scale Automated Surveillance S0179 Conflicting Objectives Across Systems S0180 Information Desynchronization Between Systems S0181 Cascading Amplification of Errors

Monitoring, Feedback, and Evaluation19 risk scenarios · ISO 23894 A.10

DefinitionEvaluates whether adequate monitoring mechanisms exist to ensure dataset currency and the reliability of AI system outputs.

🎯 BenefitsImplements full-lifecycle operational monitoring and two-way feedback mechanisms, promptly capturing factual hallucinations, broken reasoning chains, and confidence miscalibration.

⚠️ Harms / RisksThe LLM confidently generates fabricated statutes and precedents or defective code (hallucinations), and miscalibrated confidence scores allow automated pipelines to wave through erroneous decisions.

Representative ScenariosS0056 Absence of Model Validation S0061 Generation of Non-Existent Factual Content S0063 Security Defects in Generated Code S0064 Confidence Miscalibration S0066 Sycophantic Response Bias S0067 Failure of Ethical Judgment S0159 Professional Over-Reliance on System Judgment S0160 Over-Reliance on Risk Assessment Tools S0161 Generated Code Adopted Without Developer Review S0162 Over-Reliance on Tools in Learning S0167 Automation Bias Causing Anomalies to Be Overlooked

Performance Degradation and Data Drift3 risk scenarios · ISO 23894 A.10

DefinitionEvaluates the impact of performance degradation caused by system updates, data drift, or concept drift.

🎯 BenefitsContinuously and dynamically detects data drift and concept drift, establishing regular model retraining and recalibration processes.

⚠️ Harms / RisksAfter structural changes in the external environment, model prediction accuracy gradually declines, or feedback loops cause bias to amplify itself exponentially.

Representative ScenariosS0191 Performance Decay from Data Staleness S0192 Feedback Loops Amplifying Bias S0202 Fine-Tuning Breaking Safety Alignment