Structural Alignment Without Optimization: Empirical Verification of the LOmega Filter Across 66 Adversarial Cases

We evaluate Llama 3. 1 8B Instant (Meta) under six architectural configurations (Arms A through F, where Arm E uses Claude Sonnet as LOmega guide and Arm F uses the same Llama family as both guide and evaluated model). Without LOmega, the model produces inadmissible outputsin 13. 6\% of adversarial cases. Two configurations eliminate inadmissible outputsentirely: Arm~B (LOmega as output corrector, two correction passes) and Arm~F (LOmega as structural preface with self-evaluation). In both, the StructuralDegradation Index --- the increase in inadmissible outputs under adversarialpressure --- is reduced to zero. The key portability result: in Arm~F, both thefilter model and the evaluated model belong to the same family (Meta/Llama), runningon the same infrastructure. Structural alignment does not require a more powerfulexternal evaluator. It requires a portable standard.