Context-Partitioned Generic Assembly with Swarm-External Enforcement for Overcoming LLM Context Degradation

CPGA+SENTINEL is a constraint-routing architecture for large language models built on one principle: classify every in-context constraint by whether it genuinely needs the LLM's attention, and externalize everything that does not. A classification protocol (FORGE) assigns each constraint to the cheapest reliable enforcement layer (token mask, compiled checker, residual prompt, or post-generation judge); a diversity mechanism (CADG) permutes constraint order across candidates, reducing positional neglect exponentially in candidate count; and an external enforcement swarm (SENTINEL) provides concurrent per-constraint agents in four tiers, serving a dual role as production enforcement and reusable evaluation suite. Validation spans four scales: seven controlled experiments isolating each mechanism, benchmark evaluation on three public benchmarks (MOSAIC, IFBench, IFEval) with 1,081 tasks, a 50-constraint synthetic benchmark, and a three-month production deployment processing 3,000+ items against a 5,077-rule taxonomy. Public-benchmark gains are statistically significant: +6.4pp MOSAIC (Full Stack), +10.0pp IFBench (CADG+SENTINEL), and +6.2pp IFEval (Full Stack). The production deployment achieved 90%+ independent QC pass rate, ~20x per-item context reduction, and +14.3pp system-level quality improvement. This archive contains the preprint PDF, the pre-registered analysis plan, the complete experiment harness, benchmark adapters, raw JSON results, and model configuration files required to reproduce every numeric cell in the paper. Companion OpenReview submission (anonymous version) is in review at the Transactions on Machine Learning Research (TMLR).