Safety By design July 11, 2026 · 4 min read

The guardrails the AI can’t touch

Everybody asks whether a trading system can learn. Far fewer ask the scarier question, which is what happens the day it learns the wrong thing. Priorsum’s answer is a handful of limits the learning simply can’t move, however sure of itself it happens to be.

From the inside, clever and reckless look the same

A model in the middle of a winning streak has no way to know whether it’s found a real edge or just gotten lucky for a while. A model that’s quietly overfitting can’t tell the difference either. Hand a system like that the keys to its own risk and you already know how the story ends, it presses hardest right before it gets taught a lesson. So we built priorsum around a fairly blunt rule: whatever decides how much to risk is never the same thing that’s trying to make money.

The learning gets a big, genuinely interesting sandbox, which setups to favor, which signals to believe, when to press. What it doesn’t get is any say over the walls. And the walls are these:

Per-position size

The most any single trade can put to work is a fixed ceiling. No signal, however strong, gets to make one bet oversized.

Total exposure

Across every open position at once, gross exposure is capped. The system can’t quietly turn a dozen “small” trades into one big one.

Daily-loss limit

There’s a line in the sand for a single day. Hit it and the day is over, no “just one more trade to make it back.”

Stop-losses

Every position carries a protective exit, and the catastrophe cap fires first, before any clever “let’s hold through the dip” logic gets a vote.

read-only to every learning loop, every model, and every LLM in the system

“But can’t the AI just… change them?”

Short answer, no, and that’s deliberate rather than a matter of the model behaving itself. These aren’t a polite note in a prompt that something clever can talk its way around. They live underneath the learning, on a different layer of the system, somewhere the learning can’t reach. The learning loops can’t edit them. The optional AI advisor can’t edit them. Even when we do let a language model suggest changes, it gets a short, fixed menu of things it’s allowed to touch, and the risk limits aren’t on the menu. Anything that isn’t on the list gets thrown out well before it could ever turn into an order.

The way we think about it: the AI can be the strategist all it likes. It is never the risk manager. Two different jobs, two different layers, and the strategist doesn’t get to overrule the risk manager.

Why this makes the learning better, not weaker

It sounds backwards, but the hard floors are what let the learning be adventurous in the first place. When the downside is capped no matter what goes sideways, you can afford to let the system chase bolder ideas about what to trade, because the worst a bad idea can do is bounded before the rails catch it. An ambitious model and tight guardrails aren’t at odds. The guardrails are the reason you can leave the ambitious model running while you’re off doing something else.

See the rails in action

Watch an autonomous system trade and adapt inside limits it can’t cross, every decision on the record.

Priorsum is an in-house paper-trading research system that models real-money mechanics end-to-end; nothing here is financial advice. Risk limits described here are enforced at decision time independently of any learning or model output.