The problem: LLMs don't fail loudly. They drift. Mission reinterpretation, constraint erosion, format collapse. Traditional monitoring catches it after divergence.
DriftProof prevents drift through six structural invariants: - Identity Lock (role can't be redefined) - Mission Lock (purpose stays fixed) - Constraint Cage (boundaries are immutable) - Format Lock (output structure is deterministic) - Interpretive Invariance (context can't override core properties) - Operator Sovereignty (authority hierarchy enforced)
Repo includes: published specification (normative requirements), reference implementation (deploys in 5 seconds), security model, compliance checklist.
Patent: US 63/905,582 License: MIT (code) + CC-BY-SA (spec)
Looking for feedback from people building production LLM systems. Early pilots available.