Deterministic AI Control
Deterministic AI control means governing agent actions with explicit, rule-based policy that produces the same decision for the same input every time — as opposed to relying on a probabilistic model to judge whether an action is safe.
Definition
Deterministic AI control means governing agent actions with explicit, rule-based policy that produces the same decision for the same input every time — as opposed to relying on a probabilistic model to judge whether an action is safe.
Determinism as a safety property
An LLM is probabilistic by design: the same prompt can yield different outputs. That is fine for generating text, but undesirable for the question 'is this agent allowed to delete this database?' Deterministic control answers governance questions with code-defined rules — allowlists, denylists, scopes, rate budgets, and approval requirements — so the decision is reproducible, auditable, and explainable. You can point at the exact rule that allowed or blocked an action, which is what compliance and incident review require.
Where determinism belongs
The pragmatic split is to let the model plan and reason, and let deterministic policy govern the consequential actions: tool calls, data access, and side effects. Cordum applies this by evaluating each job against versioned policy in the Safety Kernel before dispatch. The agent's intelligence chooses what to attempt; deterministic control decides what is permitted. This keeps the unpredictable part of the system away from the part that touches real infrastructure.
Frequently asked questions
Why not let the model decide what is safe?
Model judgment is probabilistic and non-reproducible, so the same risky action might be allowed once and blocked the next time, with no rule you can audit. Deterministic policy gives the same answer every time and lets you prove why a decision was made.
Does deterministic control limit what agents can do?
It limits which actions execute, not what the agent can reason about. The model still plans freely; deterministic policy only constrains the consequential side effects, which is exactly the slice where reproducibility matters most.
Related reading
Govern your AI agents with Cordum
Cordum is the agent control plane: policy-before-dispatch enforcement, human approvals, and a tamper-evident audit trail for autonomous AI agents.