Operationalized Integrity Principles (OIP): A Framework for Epistemic Governance of Large Language Model Outputs in High Stakes Domains

This record contains the full release of Operationalized Integrity Principles (OIP) v1.1, a framework for improving the epistemic governance of large language model (LLM) outputs in high-stakes domains. OIP addresses failure modes such as hallucination, scope inflation, unjustified certainty, and causal over-closure by combining: explicit epistemic classification (ECG), gate-based enforcement of representational and evidentiary discipline, and deterministic evaluation via a testing framework (OIP-TF). Contents Primary paper Operationalized Integrity Principles: A Framework for Epistemic Governance of Large Language Model Outputs in High-Stakes Domains(theoretical foundation and motivation) Normative specifications OIP-Core (v1.1) — concise behavioural specification OIP Handbook (January 2026) — operational and implementation guidance Machine-enforceable artefacts OIP-JSON Standard (v1.1) — canonical enforcement schema OIP-JSON Persistent (v1.1) — persistence-oriented orchestration schema OIP-TF (v1.1) — testing and scoring framework Supplementary diagnostic materials Diagnostic self-test and illustrative evaluations demonstrating framework behaviour under controlled conditions Methodological note The supplementary evaluations included in this record are illustrative and diagnostic, not empirical validation studies. They are provided to demonstrate framework behaviour, failure-mode detection, and internal consistency under controlled prompts and model configurations. All scoring was performed using OIP-TF Release 1.1. Subsequent terminology clarifications affected narrative interpretation only and did not alter underlying evaluation arithmetic. Intended use OIP is intended for: AI governance and assurance research regulatory and compliance contexts - and other "high stakes" use audit and evaluation of LLM-based decision support systems practitioners seeking structured epistemic controls for high-stakes AI use License All materials are released under Creative Commons Attribution 4.0 (CC-BY-4.0).