The path that never ran

On Monday morning, a sold position refused to close. The system saw a stock it had sold the previous Friday, but in the database the position was still open. That stopped the entire morning run before a single order could be placed.

That was the visible problem. The real problem was deeper, and it had been there for weeks.

Fifteen orders, one trade

While digging in, a number surfaced that stopped me for a moment. Of the fifteen orders the system had ever executed, exactly one had a trade record in the books. One.

To be clear about what that means. When you buy or sell a stock, three things should be updated: the order goes to "filled", a trade lands in the books for your returns and history, and your position is opened or closed. In my case only the first happened. The order went to "filled", but the trade and the position were left behind.

The positions were still correct, but for an accidental reason: they were fetched directly from the broker, separate from this path. So everything looked fine. The dashboard was right, the positions were right, and nobody noticed the underlying bookkeeping was half done. Until the first sale, because that is when the gap became visible.

Why it never ran

The system had two ways to process a filled order. A real-time way, meant to react at the moment of the transaction. And a slower way, which periodically asks the broker which orders have been filled.

The real-time way was cleanly designed. The problem: it only listened during the half-hour window in which the system trades in the morning. But the US market opens later in the day, so transactions come in hours later, well after that listening window had already closed. The real-time processor was never present at the moment that mattered.

What remained was the slower way, and that only set the order to "filled". The trade and the position, the important part, it did not do. The primary path listened at the wrong moment, and the backup path was never finished.

What this taught me

This is the difference between "looks like it works" and "works as designed". For months the system produced plausible results. No error messages, no red flags, a dashboard that was correct. And yet an entire path did not do what it was supposed to do.

The fix itself was not even the interesting part. The backup path was made complete, so that from now on it also updates the trade and the position, and the existing backlog was cleanly caught up. What stayed with me is how long something like this can stay invisible when the symptoms happen to be masked.

And this is exactly why the system still runs on paper money and not real money. You do not want to discover this kind of thing with your own capital at stake. You want to discover it on a quiet Monday morning, with zero risk, while you still have time to do it right.

A system that produces no errors is not the same as a system that is correct. The second you have to prove. The first proves nothing.