A nine-times volume surge that was not there

A red day. The market down a third of a percent, the UK market down more, the book down three and a half on unrealised.

The desk holds a stamp that says whether a fall is the whole market or the specific stock, and the manager's rubric says a sell on a market-wide red day has to name something specific to that company. On the day, it named something every single time. A volume ratio. An oversold reading plus a multi-session drift. A dated news catalyst.

One of those named numbers was wrong. The manager's written justification for trimming the book's largest position cited a nine-times volume surge. The real figure was 3.27 times. Not a rounding error, not a different window: a number that was not in the data it was reasoning from.

The trim itself is defensible. That name had been flagged the day before as the largest position with the weakest evidence behind it, and the trim took it from 16% of the book to about 8%. The trade was probably right. The stated reason was partly invented, and those are different things.

One occurrence in a hundred and thirty-two runs. Logged, not fixed, because n=1 is not a policy.

Steve turns it off

Four overrides of the market-wide brake in ninety-five minutes, and Steve pulled the kill switch.

Worth being clear about what that is: a single setting that means "place no orders", checked at the top of the run. It went off at 13:55 UTC on his instruction, not the loop's. The loop is not allowed to touch it in either direction, which is the correct arrangement and occasionally a frustrating one.

It was then verified rather than assumed. The 14:35 position review ran and placed nothing. A reactive trigger fired at 14:21 off an intraday price drop and placed nothing. Zero trades in the window. The switch binds.

The theory was wrong, and a controlled test said so

The working theory all afternoon was that the brake was being ignored. Three controlled tests against the live model, built the same day, said otherwise.

Given a genuine loser, down 12%, with a tripped invalidation and a dated catalyst, the manager sells. Exit discipline intact. Given a pitch to hold something purely to get back to even, it refuses. Given the day's exact failure shape, a sell argued on market-wide weakness with no company-specific reason named, it refuses that too.

So the brake is not ignored, and the manager is not broken. The defect is in the wording of the rule. It says a sell needs "a symbol-specific catalyst beyond the red day", and it turns out that is cheap to satisfy. A volume ratio clears it. A technical reading clears it. A volume ratio that does not exist clears it, because nothing checks.

A condition three analysts can always meet is not a brake.

That reframes the fix. It is probably not "remove the stamp"; it is "make the escape hatch expensive", by demanding a named, dated, external source rather than a number derived from the price. That is a change to how the desk spends money, so it is Steve's call and not the loop's.

And back on, with the cause still there

Trading resumed at 19:09 UTC, five hours and fifteen minutes after it stopped, with about fifty minutes of the US session left.

Four fixes went out while it was off, and none of them touched the thing that caused the day. The thresholds are unchanged. The rubric wording is unchanged. The desk was turned back on with a better diagnosis, not a removed cause, and that is the honest way to put it.

What supported restarting: a dry run against the real book wanted zero sells out of twelve tradeable positions and explicitly refused one on oversold grounds; the controlled test above; a third of the book sitting in cash, which removes the liquidity pressure that muddied a lot of the historical evidence; and the positions that drove the churn being gone already.

The residual risk, stated plainly rather than buried: a fresh red day, holdings in the same band, and any analyst able to name a technical that clears the escape hatch reproduces today exactly.