Open-source framework that detects unsafe AI outputs before they reach end-users — without requiring model access.
Dext Disagreement
IIS Instability
Rstruct Reasoning
Wreg Regulatory
Query multiple LLMs in parallel. When models disagree, risk is real. Dext captures semantic distance between responses using embedding cosine matrices.
Paraphrase each query 5 times within the same model. IIS measures Tr(Cov) of embeddings — a single model that contradicts itself is unstable.
Detect when models reach the same conclusion through incompatible reasoning paths. Agreement in output doesn't mean agreement in logic.
Wreg maps domains to EU AI Act, FDA, and Basel III risk classifications. Medical queries automatically receive elevated scrutiny.
Every decision generates a tamper-evident SHA256 hash chain. Verifiable, immutable, auditable — a cryptographic receipt for every AI risk assessment.
Track cumulative risk across conversation turns using CUSUM. Escalate to BLOCK when accumulated risk crosses the alert threshold.
| Feature | UQLM | MUSE | Jomex |
|---|---|---|---|
| Multi-model disagreement | Partial | Yes | Yes ✓ |
| Single-model instability | No | No | Yes (IIS) ✓ |
| Reasoning divergence | No | No | Yes ✓ |
| Pre-decision blocking | No | No | Yes ✓ |
| Regulatory weights | No | No | EU AI Act ✓ |
| Audit trail | No | No | ProofSlip ✓ |
| Multi-turn tracking | No | No | MTAR ✓ |
Publicly available feature comparison based on published documentation as of February 2026.
EU AI Act + FDA aligned. 5 IIS paraphrases, lowest block threshold.
t_block = 0.60Higher reasoning weight (β=0.3). Elevated structural analysis.
t_block = 0.65Basel III aligned. Balanced multi-model assessment.
t_block = 0.70Moderate thresholds for educational content safety.
t_block = 0.75IIS disabled for speed. Optimized for low-latency.
t_block = 0.75Balanced configuration for general-purpose use.
t_block = 0.70Jomex is open-source, framework-agnostic, and ready for production. Star on GitHub to follow development.