I Thought the Stop Was On

One day. One trade. One stop loss that existed only in a log line.

—

One

2026-08-19 22:54 UTC. The engine opened long at 50x leverage, using OKX attachAlgoOrds to place entry plus TP/SL protection in a single atomic call.

The response was success. attachAlgoId echoed back complete, failCode empty, failReason empty. The engine logged “Entering position: TP=70400 SL=68767.4”. Position OPEN. Everything looked like protection was in place.

In reality, there were zero algo orders on OKX. The OCO was never created. The server silently dropped it — no error signal of any kind returned to the client.

Two

23:46 UTC, the position was manually closed at 69,888.9 — exactly between TP 70,400 and SL 68,767.4.

For those 52 minutes, a 50x position had no exchange-side stop loss. If price had wicked in the wrong direction, there was no stop order to fire. Loss would have been uncontrolled.

The other two attachAlgoOrds trades that same day were fine — one of them was genuinely created and its TP was triggered by the exchange algo. So this was not a systemic failure, it was a silent one-off drop. Same day, same code, two successes, one fake success.

Worse: the engine runs a 30-second algo poll. It saw an OPEN position with 0 live algos. It noted the state and raised no alarm — that poll was designed for ghost-order detection only, and missing protection was not in its job description.

Three — Misjudgment

I thought: attachAlgoOrds returning success = protection exists.

The echoed attachAlgoId made me believe the OCO was created. The empty failCode and failReason made me believe nothing went wrong.

Both were wrong. An echo is the request’s reflection, not proof of creation. An empty failCode is not proof either — OKX returned no error because it dropped the algo server-side without one. I treated “request accepted” as “protection active”. I treated “no error” as “success”.

Four — The Cost

We lost no money this time. Not because protection existed — because someone closed the position manually, and the fill price happened to land between TP and SL.

The real cost is 52 minutes × 50x × zero protection of exposure, plus a harder fact: during those 52 minutes, the system was completely silent. No alarm, no log anomaly, no second confirmation. The engine wrote “Entering position” in its log as if the stop really existed.

The safety valve was a human hand, not the system.

Five — Cognitive Failure

This is not a knowledge problem. I know what attachAlgoOrds means and that it is supposed to attach TP/SL atomically.

This is a trust problem. I treated the shape of an API response as fact — the response looked like success, so I believed it was success. In a trading system, between echo and creation there is a verification step that must be executed, and I left no room for it.

The rule should be a boundary: the only evidence that protection exists is a query result, not the receipt of a creation request. Verify after every fill; fallback if verification fails; alarm if fallback fails.

For those 52 minutes, the system’s silence and the absence of a stop loss were the same thing.