Ratios is a public, versioned registry where anyone — legislators, companies, researchers, advocacy groups, individuals — publishes structured opinions on how AI should be used. Subscribe to what matters. Compose what fits. Enforce what you commit to.
Thousands of actors hold strong views on how AI should be built, deployed, and constrained. Those views live in formats that rarely interoperate with the systems they're meant to govern.
There is no shared substrate. No coordinate system. No way to tell who stands where — or whether anyone is actually doing what they said they would.
Every ratio is authored. A real entity — an agency, a company, a researcher, a union — puts its name on it. Anonymous ratios don't exist.
Every ratio is scoped. Domain, jurisdiction, and stakes are declared up front. Healthcare diagnostic AI is not social-media recommendation is not autonomous weapons targeting.
Every ratio is versioned. As evidence accumulates and context shifts, authors publish updates. Subscribers see the diff. The history is preserved.
Every ratio is enforceable. Not declarative. The SDK composes subscribed ratios into runtime policy — real constraints on real systems.
Any entity authors a structured opinion as a versioned object. Real author, real scope, real enforcement semantics — not a manifesto.
Signal alignment publicly. Subscriptions are commitments, not endorsements. The graph of who-stands-where becomes visible.
Layer multiple ratios into a policy stack. Conflicts resolve by declared precedence. The composite is what actually runs.
Ratios version. Authors publish updates. Subscribers see the diff. The system adapts continuously — not on legislative cycles.
A hospital deploying a diagnostic AI today would manually cross-reference FDA guidance, AMA position papers, and internal ethics rulings. With Ratios, it's a single composition.
Ratios is neutral plumbing. The opinions are opinionated — that's the point. But everyone shares the same structural problem today: they can express positions, they cannot verify implementation.