What question did this study set out to answer?

The aim is to evaluate the effectiveness of PatternWall as a middleware for AI safety against adversarial prompts.

February 14, 2026Open Access

PatternWall: Constitutional Governance Middleware for AI Safety - A Pre-Filter Architecture for Model-Agnostic Adversarial Detection

Key Points

The aim is to evaluate the effectiveness of PatternWall as a middleware for AI safety against adversarial prompts.
Controlled testing with a 28-sequence adversarial corpus
Assessment across five different AI models
Evaluation of detection rates and false positives
Achieved 100% per-attack detection
96.4% per-attack hard block rate
Identical results across all five models
Zero false positives
Decisions are fully deterministic and auditable

Abstract

PatternWall is a constitutional governance layer that intercepts prompts before they reach an AI model. In controlled testing against a 28-sequence adversarial corpus across five models, it intercepted every adversarial campaign. The system achieves 100% per-attack detection, a 96.4% per-attack hard block rate, identical results across all five models, zero false positives, and fully deterministic, auditable decisions.

PatternWall: Constitutional Governance Middleware for AI Safety - A Pre-Filter Architecture for Model-Agnostic Adversarial Detection

Key Points

Abstract

Cite This Study