Deterministic governance for probabilistic agents
AI agents reason in probability distributions. The systems that govern them must reason in absolutes. Mixing the two is a category error.
The category error
When an AI agent decides whether to enter a position, it is reasoning in a probability distribution, over expected return, over likelihood of a regime change, over the confidence of its own forecast. That is appropriate for what the agent is being asked to do.
When the governance layer above the agent decides whether to *allow* that order, it must not reason that way. It must apply hard, deterministic rules: this position size exceeds the cap, therefore it is blocked. There is no probability here. There is a rule and there is the order.
Mixing the two, putting a probabilistic governance layer above a probabilistic agent, produces the worst of both worlds: a system that is "mostly" compliant, with no way to prove the failure mode in advance.
What deterministic governance looks like in code
ACIE's decisioning core is a few hundred lines of straight-line code that evaluates a configured policy against an order and returns one of three values: ALLOW, BLOCK, or WARN. There is no model. There is no inference. There is no learning. There is a state, a rule, and an order.
We considered, briefly, allowing operators to write rules in a Python DSL. We decided against it. The whole point of deterministic governance is that the rules cannot drift from what the operator believes they are.
A governance layer you cannot reason about in your head is not a governance layer.
What is delegated to where
Models propose. Agents decide. Governance enforces. These are three different jobs and they belong in three different parts of the system.
The views expressed are those of the author at the time of writing and may change without notice. This publication is for informational purposes only and does not constitute investment advice or a solicitation in any jurisdiction.