The Logos Made Code: Analogical Compression and the Limits of Self-Audit

Key Points

• This research explores the limitations of self-auditing in large language models (LLMs) using analogical compression.
• Utilized orthogonal probing to identify failures in a production system.
• Compared failure discovery rates between orthogonal probing and same-axis self-audit.
• Analyzed 58 documented failure classes using persistent homology to find topological structures.
• Implemented a production reference with 85 deterministic validators and 524 end-to-end tests.
• Orthogonal probing discovered approximately 80% of failures, while self-audit revealed only 20%.
• Identified 20 significant topological loops in failure classes using persistent homology.
• Demonstrated that LLMs struggle to audit their own compression mechanisms, indicating structural limits.