Abstract
Structured logging techniques are described for natural-language reasoning artifacts produced during inference in LLM-based recommendation systems. A reasoning chain may be decomposed into structured fields at multiple sensitivity levels, including non-sensitive numerical fields, low-sensitivity categorical fields, a scrubbed rationale, and a raw rationale. Sensitivity classification may include a first layer detecting explicit personal identifiers using pattern matching and a second layer detecting semantically sensitive categories using a sensitive-category dictionary. Scrubbing replaces sensitive sentences with category-level abstractions. The structured fields are routed to different storage tiers with differential access controls, including audit-logged access for restricted tiers. A retention policy engine enforces content-aware retention periods with automatic expiration, including shorter retention for explicit personal identifiers, and may support deletion requests across tiers using a pseudonymized user identifier.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Privacy-Preserving Structured Logging Architecture for Semantic Artifacts in Machine Learning Inference Systems", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10719