Vary_artifacts

Veritas is a verification engine for multimodal AI “CI/CD for generative models.” It generates copy and images with Meta’s models, then adversarially cross-checks every artifact against a declarative policy in a closed loop until it provably passes, refusing to ship anything it can’t verify. A whole domain is just a YAML policy pack, so one engine guards five real-world problems: brand safety, accessibility (WCAG), health-claim safety (FTC), robotics synthetic-training-data, and AI misinformation. It pairs probabilistic LLM critics with deterministic Z3 formal proofs and grounded SAM 3.1 vision evidence, short-circuits + caches to cut credits ~46%, can burst critics to an inference partner, surgically edits flawed images in place, and ends every run with a cryptographic receipt that can be embedded into the image as a C2PA Content Credential.