GenAI Guardrails: Who's doing this best as a service? The Scrypted Network has three modes: { Strict | Permissive | Open }. "Strict" prevents mature/explicit content by default, and in exchange has access to top, proprietary models. "Permissive" (default) has a two-pronged strategy of first attempting the work with multiple providers (in case of false positives), and falling back to rewriting the request for safety. "Open" doesn't judge content, and finds fulfillment from a decentralized network of providers, which may be more open to mature content. --- That said, to prevent *illegal* content we still have strong guardrails, like constitutional classifiers. The Catch-22? Relying on external models for judging content can, itself, be a violation of provider Terms of Service. (Like being auto-routed to OpenAI's secretive models). Who's doing this best? Anthropic showed off some really strong capabilities in their recent security hackathon. Who else?

Thought experiment: how would you decentralize the actual guardrail itself while maintaining user privacy? One idea would be treating it as a "mixing" service: like Tornado Cash for content instead of coins. You'd have multiple nodes that receive batches of content to review, and they use ZK proofs to attest to legal/illegal + content rating for each of them, but don't know the sources. Then the gateway could reject work and provide proof of "the why" without being accused of unnecessary bias/censorship. Then they'd be part of an incentivized network at the gateway layer.

Example: say I created a wrapper to your AWS credentials w/ access to Bedrock Guardrails. Then for contributing to the mixing we paid a small premium in ETH on Base, managed, maybe, by a prototype coordination service built in @eigenlayer EigenCloud. Then anyone could join in and distribute the load and content mixing. I need to measure the latency for this.