Frontier AI: thinking frameworks for high-stakes problems
In Michael S.U. Hudson’s practice, frontier AI means thinking frameworks that shape how AI reasons and works on high-stakes problems. It is applied method, not the building of foundation models.
01
What is a thinking framework for AI?
A thinking framework for AI is a written set of rules that decides how an AI system reasons, what it may do on its own and when it must stop. It sits above the model: the model supplies language and pattern recognition, and the framework supplies judgement about evidence, authority and risk.
A working framework has five parts. A constitution states what the system is for. Decision rights set three tiers: what the AI decides, what it proposes for approval and what only a person decides. A challenge duty requires the AI to question recommendations that are over engineered or rest on assumptions, including its owner’s. A truth layer names one authoritative source for each kind of data. Kill criteria define, from the first day, the evidence that would shut the system down.
02
Where it applies
Thinking frameworks matter wherever an AI output can commit money, reputation or a decision that is hard to reverse. Three settings recur in this work.
Leadership decisions. In DIOS, AI logs decisions, retrieves past context and tracks commitments, but never makes a decision. Every write to the system of record passes an eight rule standard, including verbatim checking of quoted evidence.
Analysis of people. Portraitor runs a fixed analysis prompt that reads a conversation through four lenses, and every finding carries a confidence score and its stated limitations.
Public claims in regulated industries. An evidence label marks how well each AI drafted claim is supported before it reaches the person who approves it.
03
Engagements
Frontier AI work is taken on by introduction, and each engagement starts with the framework rather than the technology. Two pieces of this work are documented as case notes: DIOS, a decision intelligence system for C-suite management with Project 54 as its first user, and Portraitor, a private by design analysis product developed with Multinomial AB and Daniel Skatov.
Each began with written rules: what the system is for, what it may do without approval, which sources it treats as truth and what evidence would end it. The build followed only once those rules were approved. Introductions are by request.