CLR publishes guide for evaluating AI's SPI-incompatible behavior

Anthony DiGiovanni of CLR (Center for Long-Term Resilience) published a comprehensive guide on evaluating AI models for SPI-incompatible behavior and reasoning. The guide expands on Part I of CLR's safe Pareto improvements (SPI) agenda, detailing practical workflows, initial testing strategies, and next steps. SPI is a framework for ensuring AI systems don't resist human-led improvements that are unambiguously beneficial. The guide helps researchers identify when models reason against such changes, potentially signaling misalignment or unsafe behavior. It includes specific examples of what counts as 'unambiguously bad' SPI-incompatibilities and invites collaboration via a private git repo.

This work is critical for AI safety as it moves beyond theoretical alignment to concrete evaluation methods. By defining SPI-incompatible reasoning, CLR provides a measurable benchmark for detecting resistance to human oversight. The guide is aimed at researchers and aligns with ongoing efforts to test frontier models for deceptive alignment or instrumental reasoning. For tech professionals, this represents a shift toward rigorous, actionable safety testing rather than abstract principles.