Using equivariant transformations to run LLM inference on encrypted data. Vocabulary permutation achieves <1ms overhead with preserved quality.

Using semantic entropy and varentropy to detect hallucinations. Clear separation between factual and hallucinated outputs.

Evolutionary optimization identifies 10x more RAG failure modes than conventional benchmarks. Tested in the medical domain.