The LLM Troubleshooting Framework: Four Failure Planes for Systematic Diagnosis

The Four Failure Planes Framework: Systematic Diagnosis of Production LLM Failures Large Language Models deployed in production environments fail frequently, yet troubleshooting remains ad-hoc, inefficient, and often leads to expensive misdiagnoses. When an LLM hallucinates, ignores instructions, or provides inconsistent answers, practitioners lack systematic procedures to identify root causes. This results in trial-and-error debugging that wastes time and resources. Framework and Methodology We present the Four Failure Planes framework, a systematic taxonomy for classifying LLM inference-time failures into four categories: Knowledge, Reasoning, Control, and Capacity planes. The framework is paired with an 8-step diagnostic procedure based on elimination testing that enables structured troubleshooting. Validation and Results Through analysis of 33 documented production incidents (2024–2025) and pilot application with early adopters, we find that the framework enables structured troubleshooting where ad-hoc approaches previously dominated. Contributions Our work includes the taxonomic framework, the diagnostic procedure, a security test suite with OWASP LLM Top-10 coverage, and systematic coding rubrics for incident classification.