The Emerald Foundry
We build the system, run it, and write the policy from what it shows.
Build first, then write the rule
The position
Most AI policy is written a long way from the systems it regulates. A rule gets drafted, published, and handed to the people who have to live under it. They are the ones who find out what it costs, what it accidentally prohibits, and whether it could have been built in the first place.
We treat that as a problem of legitimacy, not efficiency. If you want to require something of the people building AI, you should have built a system that has to meet the requirement.
So we build first. Every framework we publish arrives with a working version of the system it describes, and we say plainly what building it changed about the framework. Where the system is one that people need, we finish it and ship it as a product.
The programme
The Emerald Foundry
The Foundry works on AI safety and responsible technology: how a capable system is made answerable, how a person contests a decision a machine made about them, and what a developer can actually be held to. Every project has three parts: a working system, that system offered to anyone who needs it, and the legislative proposal it supports.
When a committee asks what a rule would cost, or whether a developer can meet it, we answer with figures from a system we run rather than an estimate. When the system turns out to be something people need on its own, it leaves the Foundry as a product.
How a piece of work moves
04 stagesFind the failure
Start from the system itself rather than the existing rulebook. What does it do, to whom, and which parts matter? Most policy inherits the last draft's assumptions. This is where we throw those out.
Build it for real
Build the system as it would have to exist under the rule we intend to propose. Not a demo: software, and where it matters hardware, that runs under real conditions with real users.
Run it as a service
Run it, and make it available to whoever needs it. That produces real figures on cost, throughput and failure, and puts a working model in the hands of the people who would have to comply.
Take it to legislators
Publish the framework with the operating record behind it, including where building the system forced the framework to change. A paper on its own can never report that.
Where the Foundry works
06 areasFrontier AI Safety
How can a powerful AI system be made answerable before it fails rather than after?
Automated Decisions
How can a person contest a decision a machine made about them, and win?
AI in Health
How can AI act for a patient rather than only on one?
Privacy and Data
How can a system prove it holds only what it needs?
Identity and Access
How can a person prove who they are without handing over everything they are?
AI and Work
How can capable machines take on the work nobody is left to do without lowering the standard it is held to?
We take on problems where the engineering and the regulation have to be worked out together. If that is what you are stuck on, put it in the subject line.