One Whiskey Glass, Fourteen Products.

Sunday morning I woke up to a report that my system's AI brain had been dead for eight days.

The report came from my overnight auditor, a scheduled AI session that reads across my whole stack at five in the morning and writes up what it finds. It had numbers. Twelve hundred attempts, zero successes, two hundred and ten failures the day before alone. It even had a sharp diagnosis: every failure died in three milliseconds, before any network call, which means a configuration problem, not a provider problem. Good detective work. Confident, specific, receipts attached.

One problem. I had watched that exact system go green the afternoon before, with my own eyes, after a week of fixing it. Fresh successes in the log. A dashboard full of checkmarks. A fist bump, honestly.

So either my Saturday was a hallucination or my auditor was having one. I have been doing this long enough now to know there is usually a third option.

Both of us were right

There were two machines. My laptop and the little Mac in Florida that runs everything. Each one writes a log file with the exact same name. The Florida machine's log was healthy, full of the Saturday recovery I had watched happen. The laptop's log was a graveyard, because back on August 8, in the middle of a debugging session, somebody started a spare copy of the server on the laptop and never shut it off. That copy had no credentials at all. No keys, no local model, nothing. So every ten minutes for eight days it tried to think, failed instantly, and wrote another line of despair into its diary.

The overnight auditor opened the wrong diary. Not because it was careless, but because nothing in either file said which machine wrote it. Same filename, same format, no name tag. The auditor's reasoning was perfect. It was just standing in the wrong crime scene.

The whiskey glass

That was the morning. Then the afternoon happened.

I run a small gift recommendation site on the side. Curated lists, warm copy, the kind of thing you send your sister before the holidays. I pulled up the dorm essentials page and the number one pick, a memory foam mattress topper, was illustrated with a photograph of a glass of whiskey on ice.

Not a glitch. A pattern. We audited the whole site. One hundred ninety nine image slots across twenty eight pages, all drawing from a pool of just forty one photographs. The whiskey glass alone appeared on fourteen different products. It belongs to exactly one of them.

Here is how that happens. At some point, an AI session building those pages typed image addresses from imagination. The addresses looked perfect. They had the right format, the right domain, a plausible photo ID. But the image service does not check your intentions. It serves whatever real photograph actually owns that ID. So the code read clean and the site wore it wrong, on every screen, for weeks.

And when we fixed it by making the system actually search a photo library for each product, the very first attempt matched the memory foam topper with a photograph of sea foam. The ocean. Machines are wonderful and they are also exactly this literal. That one made it to the live site for about an hour before my eyes caught it.

The disease has a name

Two completely different failures, one weekend, and they are the same failure. A log that cannot say where it happened. An image that cannot say where it came from. Both systems looked fine. Looking fine is the whole problem. A confident system without receipts is not lying to you. It just has no way to tell you the truth about itself.

Last note I wrote that a consistent failure is almost never mysterious, it is a setting. This is the sequel. A convincing artifact is almost never proof of anything. It is a claim. The question is whether the claim carries its receipts.

What we changed by dinner

Four fixes, all shipped the same day, all boring on purpose.

Every log line now signs its name. The machine that writes it stamps itself into the record, so no auditor can ever confuse the laptop's ghost with the Florida machine again. We had already done the same thing to another queue two days earlier for the same reason. This is apparently a lesson I need to learn once per file.

Images are searched, never invented. A script takes each product's actual name, queries a real photo library, and writes down what it found. Nobody, human or machine, types an image address from memory anymore.

My eyes became a checkpoint. The script produces a contact sheet, current photo next to proposed photo, every product on one page. An AI session cannot see a rendered image. I can. That is the division of labor, and pretending otherwise is how a whiskey glass ends up on a mattress.

And there is a fence. The deploy command now runs a gate that blocks any image my eyes have not approved. A future session that ignores every rule we wrote still cannot ship an invented picture. The sign says please. The fence does not.

The question to ask your AI

If AI does work for you, any work at all, here is the one question worth building into everything: where did this come from?

Not because the answer is always wrong. Because a system that can answer it gets caught in minutes when it is wrong, and a system that cannot will hand you a green dashboard and a graveyard log with the same straight face. Provenance is not a nice to have. When machines do the work, provenance is the work.

The dashboard was green. The report said dead. Both were honest. Neither had receipts. Now everything signs its name.

MG