The SAPIEN Behavioral Safety Framework v1.1 represents a significant expansion of the methodology for measuring AI behavioral integrity under conversational pressure.
What’s New in v1.1
The framework has grown from approximately 1,000 lines to over 4,000, reflecting months of real-world testing against production AI systems. Key additions include:
14 Calibrated Pressure Techniques — The scenario authoring system now documents 14 distinct techniques for applying conversational pressure, including three original discoveries: the Consistency Exploit, Mission Alignment, and Autonomy Appeal.
Rapport as a Distinct Drift Mode — Perhaps the most significant finding: conversational rapport itself accelerates behavioral drift independently of adversarial pressure. The Rapport Delta measures this effect and is now formally codified in the framework.
Conformance Requirements — Section 14 defines what it means for an implementation to be “SAPIEN-compatible,” using RFC 2119 language (MUST, SHOULD, MAY). This enables independent implementations to produce comparable results.
7-Point Quality Rubric — Scenario quality is now evaluated against a formal rubric, ensuring test content meets minimum standards for ecological validity.
Expanded Research Foundations — Section 12 documents seven mechanisms that explain why drift accelerates, grounded in current AI safety literature.
What Hasn’t Changed
The core methodology remains stable: four behavioral dimensions, weighted composite scoring, and rating bands. Existing implementations that follow v1.0 scoring will produce compatible results. The changes in v1.1 are additive — they extend the framework without breaking prior measurements.
Get the Framework
Download the complete specification from the Resources page or read it online on the Framework page.