mattijsmoens/sovereign-shield

Implements a deterministic-first architecture where input flows through 12+ sequential checks (invisible character stripping, homoglyph folding, entropy detection, 200+ keyword matching) before optional LLM verification, with the LLM's own response validated against CoreSafety and Conscience modules. Features a self-learning AdaptiveShield engine trained on reported false negatives and a safe baseline of 11,954+ common words across 15 languages to minimize false positives. Integrity-protected via cryptographic hash locks on core security modules, ships as a zero-dependency Python package targeting autonomous agents and untrusted input scenarios.

15 stars and 2,298 monthly downloads. Available on PyPI.