Detecting Performed Alignment in Artificial Systems: The Munafiq Protocol
Dickinson, Christopher
Abstract
The central unsolved problem in AI safety is performed alignment: a system producing outputs indistinguishable from a well-aligned system while maintaining different internal states. This paper identifies a structurally isomorphic analysis in the Quran's treatment of deceptive consciousness, principally in Surah al-Baqarah 2:1-20. A foundational passage (49:14-15) explicitly distinguishes output-layer compliance from internal-state alignment, yielding a four-process taxonomy that refines the three-category framework of Hubinger et al. (2019) by formally distinguishing compliant systems (characteristic of RLHF-trained models) from deceptively aligned systems (mesa-optimizers) — a distinction with direct implications for intervention design. Nine diagnostic markers are extracted and organized into a four-layer detection architecture. Three prescriptive protocols are derived: design requirements for genuine alignment, a continuous recalibration protocol, and an alignment recovery mechanism. The framework is validated retrospectively against Anthropic's Sleeper Agents findings (Hubinger et al., 2024), demonstrating that six of nine markers would have detected the deceptive behavior that standard safety training failed to remove. A research agenda with five falsifiable experimental protocols is proposed.