Launch prototype: this page is a structured draft and requires legal, security and organisational review before production use.
Core principles
- Human accountability cannot be delegated to an AI system.
- Recommendations should show evidence, confidence and important assumptions.
- Models should operate within defined tools, data and action boundaries.
- High-impact actions should require explicit approval.
- Failures, retries and changes of model or provider should be observable.
- Private data should not be used for model training without explicit lawful agreement.
Verifiable AI
COSRYX AI products are built on the AI Sovereignty Oracle — every model version is registered, every inference is receipted, every model update is registry-signed. Theft detection, poisoning cascade detection and carbon tracking are built in, not bolted on.
Revoking a model version flags every past inference on verify — the poisoning cascade guarantee. A poisoned model cannot hide its history.
Per-inference carbon tracking with tamper detection. Verifier recomputation catches greenwashing claims.
C2PA-compatible provenance manifest export for standards-aligned content authenticity.
This is not a policy promise; it is tested code (58 tests, 0 fail, 0 CVEs).
Research limitations
Model outputs can be incomplete, incorrect or sensitive to inputs. Simulation results do not establish operational safety.
Evaluation
Each use case should have quality, safety, latency, cost, robustness and human-review criteria before pilot or deployment.
Last prototype update: 23 July 2026.