Decision Reliability for Real-Time Release in Pharmaceutical Manufacturing: Modeling Sensor Drift, Sampling Error, and Time-to-Release Under Quality Constraints

Abstract

This article develops an engineering-oriented framework that treats real-time release as an end-to-end decision pipeline, quantifying how uncertainty propagates through measurement, inference, threshold governance, verification logic, and disposition actions to determine probability of false release, probability of false rejection, time-to-release, nuisance hold rate, and expected cost of quality under batch and continuous production contexts. A scenario-based quantitative study is presented for a representative solid oral dose process where near-infrared spectroscopy and inline weight/density sensing are used to infer critical quality attributes, and four operational architectures are compared: baseline fixed-threshold PAT, increased sampling without governance, model-based inference with limited drift management, and a governance-optimized two-tier architecture that combines nuisance-constrained thresholds, drift-aware verification, adaptive sampling, and staged disposition. Results show that quality risk is dominated by tail events in which drift or sampling bias produces false stability and delays intervention, while operational sustainability is dominated by the rate and duration of nuisance holds, and that a governed two-tier approach provides the best cost–risk balance by shrinking the tail of time-to-decision while controlling nuisance actions through explicit false-alarm constraints and verification pathways. The paper provides copy-ready results tables and full prompts for data-driven figures suitable for Techne submission and adaptation to site-specific datasets.