Most enterprise data stacks now make data accessible. They do not make it safe to interpret. That distinction is where many modern analytics failures live. In production, the expensive mistakes are often not broken pipelines or failed SQL queries. They are plausible answers generated from incomplete operational knowledge. The SQL runs. The dashboard renders. The meeting happens. The decision is wrong. This paper introduces the Living Knowledge Layer (LKL), a governed, versioned, and injectable repository of operational knowledge for enterprise analytics. The LKL captures rules, contexts, traps, mappings, join patterns, exceptions, precedence rules, and cross-domain relationships that are usually stored in people, meetings, wiki pages, tickets, and undocumented analyst habits. The LKL is designed for LLM-driven analytics, but it is not just prompt context. A mature LKL acts as a control plane. Validated entries are retrieved and injected before an LLM generates SQL or explanations. Mandatory entries can also compile into validators, forbidden-pattern checks, table-selection constraints, refusal conditions, and audit trails. The semantic layer remains necessary. The claim of this paper is narrower: semantic layers do not, by themselves, capture the full operational knowledge required to prevent silent analytical failures. The LKL complements the semantic layer by making operational knowledge explicit, reviewable, testable, and applicable at runtime. This paper reports a twelve-month industrial deployment that produced 235 knowledge entries across eight functional domains, and proposes an evaluation protocol against schema-only, semantic-layer-only, raw-documentation RAG, LKL, and LKL-plus-control baselines.
Olivier Nachba (Tue,) studied this question.