EDIP: Edge Deterministic Ingestion Protocol — Closed-Loop Relational Constraint Reflection for Autonomous Document Ingestion on Quantized Edge Models

Enterprise document ingestion pipelines face severe reliability and compliance bottlenecks when deploying generative artificial intelligence at the physical edge. Small Language Models (1B–3B parameters) deployed on local, air-gapped hardware frequently exhibit stochastic hallucinations, type coercion failures, and relational key inconsistencies. Concurrently, heightened data sovereignty directives. exemplified by official European and Swedish public sector bans prohibiting civil servants from routing administrative workflows through third-party cloud AI APIs. have rendered remote inference architectures legally non-viable for sensitive operational records. This paper presents the Edge Deterministic Ingestion Protocol (EDIP), an autonomous middleware architecture that disciplines probabilistic SLMs using native relational database integrity constraints. By converting runtime SQLite IntegrityError exceptions. specifically Foreign Key violations, CHECK constraints, and UNIQUE collisions. into an iterative error-reflection penalty loop, EDIP forces local SLMs to self-correct unstructured payloads prior to transactional commits. Operating completely offline on commodity CPU hardware without cloud APIs or heavy runtime wrappers, EDIP combines dynamic schema introspection, in-process Write-Ahead Logging (WAL) concurrency, and a tri-folder isolation architecture with a Dead-Letter Queue (DLQ). Empirical stress-testing against adversarial documents demonstrates robust self-recovery from structural hallucinations, foreign key re-alignment, prompt injection defense, and idempotent duplicate handling.