68% of Agent Omissions Come From Deterministic Middleware, Not the Model: a 75,476-Trial Sweep of Air-Gapped LLM Pipelines

arXiv 2607.22448·medium signalXBlueskyLinkedInCopy link

'Where Facts Go Missing' (arXiv 2607.22448, July 24) targets on-prem and air-gapped deployments running quantized 4–8B models via llama.cpp or vLLM behind tool servers, where the dominant failure is omission — an agent reading 20 of 400 records and reporting 'no anomalies.' A 75,476-trial sweep across five models and two engines found a pooled omission rate of 0.62, and 68% of it originates in deterministic middleware layers (L0–L3: ingestion, chunking, retrieval plumbing) rather than in model behavior. The paper contributes a nine-layer omission taxonomy (L0–L8), an attribution method separating deterministic from behavioral layers via controlled ablation and logit decomposition, and an open cross-architecture harness — relocating where operators should actually intervene.