Deterministic Controller Layer (DCL): A Deterministic Runtime Decision Control Layer for ML Pipelines

Modern machine learning systems are optimized for prediction accuracy but are not inherently designed for safe and auditable decision-making in production environments. Most models always produce an output, even under high uncertainty or when input data lies outside the training distribution. This work introduces the concept of the Deterministic Controller Layer (DCL) — a lightweight runtime system component that performs deterministic decision control using model uncertainty signals, without modifying or retraining the underlying model. DCL enables runtime risk control, deterministic auditability, and regulatory-aligned decision enforcement in production ML pipelines. The layer operates using canonical uncertainty metrics such as probability margins, entropy, and distribution mass, and produces stable decision classes including automated acceptance, risk escalation, or out-of-distribution blocking. The proposed architecture is model-agnostic, computationally lightweight, and designed for integration into existing ML systems without infrastructure changes. DCL addresses a critical gap between probabilistic prediction models and deterministic production decision requirements.