Applying Safety Instrumented System Principles from Industrial Process Control to Deployed Large Language Models

Abstract

Deployed large language models operate as open-loop, feed-forward systems: weights are frozen at release, inference contains no feedback path, no comparator, and no mechanism for runtime self-correction. Corrections supplied by users do not persist, so the same error recurs in every new session. This paper presents the Closed-Loop Runtime Learning (CLRL) architecture: five functional layers — instrumentation, deviation detection, validation, actuation, and historian — that close the loop around a frozen-weight inference engine under the discipline industrial process control applies to safety instrumented systems. The validation layer is decomposed into an immutable Safety Instrumented System (L3a), which carries the sole SIL claim and never learns, and an adaptive validator (L3b), which may learn under L3a's enforced limits — a direct transposition of the IEC 61511 BPCS/SIS split. Knowledge is stratified into three tiers with distinct governance: immutable Tenets earned from incidents under CAPA discipline, committed knowledge with domain-specific half-lives, and ephemeral session context. Update rates are bounded on five independent axes. Applying the architecture's own common-cause discipline to the method that produced it yields the paper's central finding: the governance scaffolding is authored and audited by LLMs of the same actor class as the controlled entity (hazard H-26), and the sole node outside that common mode — the human principal who recognises which check matters — is singular, so the whole-method common-cause factor β approaches unity (hazard H-25). The specification is therefore published deliberately incomplete, because external review by recognisers of a distinct actor class is the only action that lowers the whole-method β. A local demonstrator (Claude inference, GPT-4.1 validation, deterministic gate logic) is reported as a methodology demonstration — explicitly not a safety-function demonstration.