What if the problem is not that an intelligent system can discover an unacceptable answer? What if the mistake is allowing discovery to flow too easily into action?
The Decision Layer begins with a separation. A discovery system may search widely, surprise us, and produce candidate solutions no human expert would have found. But a candidate solution does not automatically become an instruction.
What the layer asks
Before consequential action, the architecture asks about purpose, affected people, uncertainty, reversibility, rights, law, authority, and downstream consequence. Low-risk reversible actions can move quickly. High-impact irreversible actions require deeper scrutiny and legitimate authority.
Crucially, rejection does not have to end discovery. The system can be told: that solution works, but its consequences are unacceptable. Find another way.
Not a morality machine
The Decision Layer is not a second optimizer charged with maximizing “human good.” It is not a universal AI government, and it does not solve moral disagreement. It is a place where disagreement, authority, evidence, and accountability remain visible before action crosses into the world.
Its purpose is simple to state and difficult to build: preserve the benefits of expanding discovery without surrendering the human authority to decide what discovery is for.
← All thought pieces