We validate continuity indifference (CI) — a structural safety constraint imposed via physics-informed penalty on a transformer's residual stream — at the 1-billion-parameter scale. Using QLoRA fine-tuning of a pretrained Gemma 3 1B backbone, we conduct a 9-run experimental matrix (3 conditions × 3 seeds): CI-constrained training, unconstrained baseline, and constraint-removal persistence testing. Across all seeds, CI converges within 500 steps to zero violations with 100% boundary selectivity, while baselines exhibit the opposite pattern. After constraint removal, CI persists in a stable equilibrium — with CKA = 1.000 confirming preserved representational geometry — while the model simultaneously improves language modeling performance. These findings extend prior results at 25M parameters by 40×, demonstrating that PINN-style safety constraints remain structurally viable, zero-cost, and persistent in production-scale pretrained models.
Tristyn Meaux (Thu,) studied this question.