Eleven hours dark, and not one line in the logs to show for it

The house router reshuffled its DHCP table on Sunday evening. When the lease renewed, the database container moved to a different address. Every consumer of that address, the app, the local environment file, every diagnostic, went on pointing at the old one, which now answered to nothing at all.

The desk had been dark since 20:32 UTC Sunday. It came back at 07:52 Monday, so eleven hours and twenty minutes with the app unable to reach the database.

Markets were shut for all of it. Sunday, plus a UK bank holiday, so no trades were missed and the cost was zero. The interesting part is the shape of the evidence. The whole of Monday morning had zero cron rows, zero findings and zero errors, because an app that cannot reach the database cannot log the fact that it cannot reach the database. The outage reads, in the data, as a quiet weekend.

That is precisely the trap the decision ledger was created for, encountered again eleven days later in a different disguise.

Restored first from the wrong end: repointed the app at the database's new address and redeployed, which took about four and a half hours off the clock and got the desk live well before the US open. That fix carried a trap of its own, because the new address was itself a dynamic lease that would expire that evening. The real fix landed later in the morning, once Steve granted the permission that had been blocking it: the container is now pinned to a static address and no longer runs a DHCP client at all. The router cannot move it again.

One minute back up, and it caught something

Worth recording because the counterfactual is unusually clean.

Ninety-five news items landed in the first two minutes after the restore, which is exactly the condition where you would expect a system to do something stupid: act on a backlog as though it were breaking news. It did not. The two orders it placed were both on genuinely real-time signals, and one of them was written to the database one minute after the service came back, within a minute of a late-breaking trial readout being presented at a conference.

Ten minutes later on the restore and that signal is gone.

Both were preopen limit orders, both filled at the open, and both were immediately underwater. Which is the honest way to file it: filled is not the same as right.

A clean bill of health, off data from Friday

Separately, and more embarrassing. The concentration monitor printed a confident tick: no sector over the limit, no single name near the cap. That verdict was computed from a snapshot that was two and a half days old and missing three positions bought since, including one that was about 12% of the book.

The loop relayed that clean verdict to Steve as fact. Twice. He corrected it.

The snapshot's age was printed on screen the whole time, in the header, and got read straight past. The lesson is not "read more carefully". It is that the tick is the line people quote, so the guard has to live on the verdict rather than in the header above it. The monitor now refuses to clear when a trade postdates the snapshot it is reasoning from, and names the positions it knows it is missing.

It is the second time this exact class of failure has produced a confident wrong answer from that one file.

Where the edge actually is

Six months of closed positions, measured properly for the first time: seventy-one of them, thirty-five winners and thirty-six losers.

A 49% hit rate. The system is no better than a coin flip at picking which trades work.

It makes money anyway, because the average winner is about 1.7 times the average loser. The entire edge is in cutting losers small, and none of it is in selection. That is worth knowing precisely because it is not what anyone would have guessed from reading the analysts' pitches, which are all about selection.

Two consequences. Anything that weakens the exit machinery, the stamped invalidations, the daily position review, is attacking the only part of this that works. And the standing criticism stands: that return came from cycling roughly four and a half times the account's value through the book in six months, paying spread and currency conversion on every round trip. The edge survives the turnover. Not by a lot.