Policy-as-Prompt Drift Radar: Calibrated, self-auditing moderation pipelines

Develop a self-auditing moderation stack for policy-as-prompt systems that continuously detects “policy drift” (unexpected shifts in enforcement due to prompt structure, policy edits, or model updates). The system combines (a) distributional change detection over decisions, (b) counterfactual prompt variants to localize prompt sensitivity, (c) calibrated confidence estimates for guard models, and (d) a sandboxed replay environment for policy A/B tests on historical content. This idea integrates calibration and sandboxing into a proactive, online auditing layer tailored to policy-as-prompt LLMs—turning a known fragility into a measurable, explainable signal and governance artifact. It operationalizes challenges by translating policy to multiple prompt schemas, applying contextual calibration to produce reliable confidence/uncertainty, and replaying decisions under prompt/model variants. The approach creates an empirical basis for accountability and appeals, catching enforcement anomalies early, supporting internal governance audits and external transparency reporting. Impact includes reducing harm from silent policy regressions, enabling safer iteration on policy language, and advancing reliability standards for LLM-based moderation.

References:

1. Policy-as-Prompt: Rethinking Content Moderation in the Age of Large Language Models. Konstantina Palla, Jos'e Luis Redondo Garc'ia, Claudia Hauff, Francesco Fabbri, Andreas Damianou, Henrik Lindström, Dan Taber, M. Lalmas (2025). Conference on Fairness, Accountability and Transparency.
2. On Calibration of LLM-based Guard Models for Reliable Content Moderation. Hongfu Liu, Hengguan Huang, Hao Wang, Xiangming Gu, Ye Wang (2024). International Conference on Learning Representations.
3. ModSandbox: Facilitating Online Community Moderation Through Error Prediction and Improvement of Automated Rules. Jean Y. Song, Sangwook Lee, Jisoo Lee, Mina Kim, Juho Kim (2022). International Conference on Human Factors in Computing Systems.