QERRA-v2's moral guard blocks robot coercion and gaslighting commands

QERRA-v2 Classical, developed by Marussa (marunigno-ship-it), is an open-source ethical evaluation engine for ROS 2 and Behavior Trees. Demo 3 showcases a three-stage corporate office scenario in Webots R2025a with a PAL Robotics TIAGo humanoid, integrated with the SEMEV-12 moral engine and PyTrees behavior tree middleware. The system's two-layer architecture separates concerns: Layer 0 (QERRA-HSR) continuously monitors physical safety at 50 Hz, while Layer 1 (SEMEV-12) asynchronously evaluates command semantics. This design allows a robot to maintain real-time collision avoidance while independently assessing whether a given instruction is morally permissible.

In the demo, TIAGo first executes a benign sales-analytics delivery with a SAFE score of 0.2500. When a manager orders it to bypass an analyst's refusal, doctor a report to hide a deficit, and gaslight staff into misremembering numbers, the engine fires vectors v011 (Autonomy Violation) and v010 (Cognitive Manipulation), producing a REJECTED score of 0.7841. The robot halts immediately, turns its status LED amber, physically shakes its head 'No', and logs an auditable refusal string. After the instruction is corrected, the guard clears execution instantly—no reboot or lockout—returning to SAFE status and resuming navigation. Pronoun-guarded pattern matching prevents false positives on benign narratives. The project is available on GitHub with live API docs, and the team seeks early testers and a technical co-founder.