Author here.

This is a minimal proof-of-concept of an invariant semantic layer.

It does not replace embedding models. It sits on top of them and learns to keep semantically equivalent inputs close under noise and paraphrase.

noise_test.py shows a simple stability benchmark.

Happy to answer technical questions.