AI Risk Management Tool: Built Dumber On Purpose

Everyone building AI trading tools wants a smarter one. Faster signals. Sharper predictions. A model that "learns" your account and adjusts on the fly.

We went the other direction.

When Glenn and I (Reid) sat down with Claude to build our own position-sizing risk tool, we gave it one job. Get dumber, not smarter. It's allowed to do less. It's never allowed to do more.

That sounds backwards until you remember what this tool touches. Money. Real risk. The kind of decision that, if it's wrong, doesn't just cost you a bad trade. It costs you the account.

When AI touches anything near your capital, "smart" is the wrong goal. "Constrained" is.

What Does "Constrain the Tool" Actually Mean?

Before we wrote a single line of code, we wrote a sentence.

"This tool may only ever REDUCE position size. It may never increase it."

We had Claude repeat that rule back to us before anything else happened. Then every design decision after that got checked against that one line. Every feature. Every edge case. Every "what if the data looks weird here."

If a feature couldn't survive that check, it didn't make it into the tool.

Why bother? Because a tool that can only say "less, or the same" can't blow up your account, even if it's buggy somewhere else in the code.

It can be wrong. It can be overly cautious. It can annoy you on a day when you actually had a great setup. What it cannot do is talk you into oversizing a position because of a bug nobody caught yet.

That's the REPs framework in practice. Risk management goes first, before the tool ever gets to weigh in on your edge or your psychology. A risk tool earns trust by protecting you on your worst day, not by promising you a bigger win on your best one.

What Should the Tool Do When It Doesn't Know?

Here's the part most builders skip. What happens when the tool is missing data, or confused, or just isn't sure?

Most tools guess. They fill the blank with yesterday's number, or a default, and keep moving like nothing happened. That's the failure mode that actually hurts you. Not the tool being wrong. The tool being wrong confidently.

We built ours to fail closed and alert loudly. If the trade log won't load, or the drawdown number looks stale, the tool doesn't shrug and reuse yesterday's figure. It says: "NO SIZE RECOMMENDATION, data problem." Loud. Visible. Impossible to miss.

A wrong answer delivered with confidence is the most expensive kind of wrong answer.

If you're pulling stats from TradeZella into a sizing model, or pulling price data from TradingView or TrendSpider into a signal, you carry the same exposure. Any one of those feeds can hiccup.

The question isn't whether your data will glitch someday. It's whether your tool tells you when it does, or just hands you a number like nothing's wrong.

Is "Designed" the Same Thing as "Deployed"?

We learned this one the hard way, on a completely different system.

We'd built a safeguard into the connection layer of our trading platform. The part that watches for a dropped feed or a broken link between our charts and our execution. Good design, on paper. It sat in a folder for a month.

The problem it was built to catch happened daily during that month. The safeguard never fired.

Not because the logic was wrong. Because nobody had actually proven it was switched on. It was designed. It was never confirmed deployed. Those are two different states, and the gap between them cost us real time we didn't need to lose.

That one stung enough to change how we ship anything now, tool or trade plan.

How Do You Know a Safeguard Is Actually On?

We run what we call a kill test. Before we trust any safeguard, in the risk tool, in the platform, in a strategy's stop logic, we force the exact failure it's supposed to catch. On purpose. Then we watch to see if it actually responds.

Not "does the code look right." Does it do the thing, right now, when the bad thing happens.

That's a small habit that catches a big blind spot. It's easy to write a safeguard, feel good about it, and assume writing it is the same as it working. It isn't.

The only proof is watching it fire.

This is where psychology meets process. It's tempting to build something once and move on. Same instinct that makes a trader skip journaling once they've had a good month. Confidence without confirmation is just a story you're telling yourself.

We talked through more of this build on Edge Up Podcast Episode 077, "Using Claude and AI in Trading," if you want the fuller version.

The Real Lesson Here

A smarter tool sounds better in a pitch. A dumber tool is the one still working for you six months from now, on the day you actually need it. One that can only say no. One that fails loud instead of quiet. One that gets kill-tested before you trust it.

That's not a limitation. That's the point.

If you're using AI anywhere near your risk management, don't ask it to be impressive. Ask it what it can't do. Then write that rule down before you write anything else.

We cover the mindset side of this same idea in Trading Risk Management Strategy: The Psychology Edge, and if this is new territory for you, Process Over Profits is a good next read. Same principle, different angle. Patience is key here, same as it is in a trade.

Mahalo for reading and trade well! Glenn & Reid | Hawai’i Trading Academy