Paraphrastic Resistance Index (IRP) vs. Typed Decision System

This paper reports a direct ground-truth comparison between a standalone Index of Paraphrastic Resistance gate and a typed decision supervisor in a controlled three-node invoicing chain. Thirty synthetic customer records passed through three transmission modalities, producing 90 condition-level trajectories. Two structured modalities preserved the information required for verification, whereas an intermediate prose rewrite removed material data in all 30 affected trajectories. The IRP gate produced 30 true positives, 59 true negatives, one false positive, and no false negatives, corresponding to 98.9% accuracy, 100% failure recall, and 98.3% specificity. The typed supervisor produced no true positives, 60 true negatives, no false positives, and 30 false negatives, corresponding to 66.7% accuracy, 0% failure recall, and 100% specificity. A separate same-environment latency measurement found that the five-component IRP procedure added a mean of 977 ms per text, a 29% overhead relative to a typed decision. This result establishes a narrow comparative advantage: within the tested domain, architecture, and failure criterion, IRP was the safer standalone gate for detecting semantic and structural degradation. It does not establish universal superiority over typed systems. A complementary severity study further showed that IRP does not replace exact deterministic validation when a complete and inexpensive rule is available. The evidence supports proportional deployment based on authority, reversibility, transformation depth, and the cost of undetected failure.