The determinism boundary: governing AI in regulated software

Regulated software runs on guarantees. In healthcare, finance, and other audited domains, “we believe it behaves correctly” isn’t an acceptable answer — you need to be able to prove that certain things always happen and others never do. AI models, by their nature, don’t offer guarantees. They offer probabilities. A model that does the right thing 99.9% of the time is remarkable, and still unusable as the sole basis for a guarantee, because the 0.1% is exactly what an auditor, a regulator, or an incident review will ask about.

The resolution isn’t to make the model deterministic. It’s to wrap it in something that is. We call that the determinism boundary.

A deterministic shell around a probabilistic core

The determinism boundary is the deterministic layer of policy, sandboxing, approvals, and audit that surrounds a non-deterministic AI model. The model can be as creative and probabilistic as it likes inside the boundary. What crosses the boundary — what actually executes, changes data, or ships — is governed by deterministic rules that do not depend on the model behaving.

It works because the guarantees live in the shell, not the core:

• Policy defines what is permitted, in code, independent of the model.
• Sandboxing ensures the model can only act within those permitted bounds.
• Approvals put a human or a deterministic gate in front of anything that matters.
• Audit records what happened in a form you can hand to a regulator.

Enforced, not hoped

This is the distinction that matters in regulated work. A behavioral guarantee you hope the model keeps is not a guarantee — it’s a wish with good odds. A guarantee the system enforces holds regardless of what the model does on any given run, because the model was never the thing being trusted. The boundary was.

That reframes the entire compliance conversation. You stop trying to certify a probabilistic system as if it were deterministic, and start certifying the deterministic boundary that contains it. The model’s output becomes a proposal; the boundary decides what’s allowed to become reality.

Why this unlocks AI in serious domains

Teams in regulated industries often assume AI-assisted delivery is off the table — too unpredictable to defend. The determinism boundary is what makes it defensible. It lets you capture the model’s speed and capability while keeping the behavioral guarantees, audit trails, and human accountability that the domain requires. Cortex is built around this principle: the intelligence is probabilistic, but what ships passes through a boundary that isn’t.

In regulated software, you don’t get to trust the model. You get to trust the boundary you put around it.