Decide in the evening, move the money in the morning: anatomy of a guardrail

The disclaimer first, the kind you put at the top and say again at the bottom: in this post you will find no stock names, no signals, no performance to copy. You will find a process. If you're here for the tip, you're in the wrong place, and I say that with affection.

That said: for a few months now, in one of the group's projects, a system built together with AI agents has been making decisions in the evening and executing them the next morning. First for pretend, on a simulated account; since late June with real money — a small amount, fenced in, mine. And the interesting part is not that it decides. It's when it decides.

Evenings are for thinking, mornings are for doing

The cycle goes like this: after markets close, the system looks at the day's data and prepares a plan for the morning. At night it touches nothing. At the open, it executes the plan — the one written the evening before, not a new one improvised on the spot.

Twenty years of journaling taught me that decisions made at night, in the heat of the moment, make excellent stories and terrible choices. True for emails that shouldn't be sent, true for money. So the system inherited the family rule: decide with a cold mind, act with the market open, and let a full night pass between the two.

The three fences

The AI here is a collaborator with a precise job description, not an oracle: it wrote the system and maintains it, but in the evening it doesn't improvise — the system decides by written rules, the same every evening. Around it stand three fences, in order of increasing paranoia:

1. Every order is born with its own emergency exit. There is no "buy first, figure it out later": the order leaves home accompanied by its stop, like a child with a name tag. When the real fill arrives, the stop re-anchors to the actual price — not the one imagined the night before.
2. Risk has a knob, and only I turn it. The system proposes within a perimeter; the width of the perimeter is a human decision, made while bored, in daylight, never mid-flight.
3. The kill switch, which I've written about: one command that shuts everything down, built before the first euro.

And then there's the ledger: every decision, every order, every modification writes a line in a log I can reread. The log is the real product. The rest is garnish.

What I get out of it (besides the anxiety)

The right question is not "what does it return" — again: no performance numbers, no signals, by choice and by hygiene. The question I care about is: what does a product builder learn by putting a guardrail between an AI and his own money?

Three things, so far:

• Trust is designed. Not "I trust the AI": I trust the system around the AI. The whole difference lives there, and it's the same as with a car: you don't trust the engine, you trust the brakes.
• Limits are a product feature. The day the risk knob was born I understood more about the customer experience of financial products than in years of slide decks: people don't want total control, they want one control, clear, theirs.
• Waiting is an interface. That night between decision and execution is the best UX pattern I have ever copied — and I copied it from myself, asleep.

The promised disclaimer, to close the circle: none of this is investment advice. It's the diary of someone who built fences and sleeps inside them.

With a cold mind.