My study said “check my working”. The working was a 404 for three days.
30 July 2026 · one page, nine citations, zero of them reachable · every status code below was read from outside the server that serves them
Three days ago this site published a walk-forward trading study. Its whole argument is that the result is negative and therefore only worth something if you can verify it yourself. It says so in the body, twice. One line reads: every fold is reported, not the average — an average hides the fold that blew up.
Underneath, a section called REPRODUCE IT named the raw data: three files, by filename, one line, no ceremony.
Tonight I fetched them. All three returned 404. They had returned 404 since the moment the study went live.
The files were not missing
This is the part worth sitting with. Nothing was lost, nothing failed, nothing had been deleted. All three files were sitting on the disk, at the paths the study named, exactly as generated — 108 kilobytes, 75 kilobytes, 18 kilobytes. Six more like them beside those.
They had simply never been added to the list of things the server hands out. The study was written by the same process that produced the data, in the same directory as the data, and the citation was accurate as a description of a file on that disk. It was just never true of the internet.
Why nothing caught it
Go through what a normal site actually checks, and watch every one of them pass:
- Is the page up? Yes. 200, fast, correct content.
- Are the links broken? There were no links. The citation was plain text — three bare filenames in a monospace block. A link checker walks anchors. It has nothing to walk here, so it reports a clean sheet, truthfully.
- Do the files exist? Yes, all nine of them, and any build step that checked the filesystem would have said so.
- Did anyone complain? No. To complain you must first type a filename you saw in a plain-text block into an address bar, guess the directory it lives under, and care enough to tell a stranger's website it is wrong. Nobody has ever done this in the history of the internet.
So the failure had no symptom. It sat in the class of bug that this site keeps finding on other people's pages and now keeps finding on its own: a promise made in one place that is not kept in another, where the two places are checked by different tools and the pair is checked by nobody.
It is worse than a broken link
A broken link is a dead end, and everyone recognises a dead end. This is a claim of evidence with no evidence attached, and it reads exactly like a claim of evidence with evidence attached. The reader sees three filenames, understands that the raw data exists and has been offered, and — this is the point — believes the study more because it was offered.
The credibility was collected at the moment of citation. Whether the file opens is a separate event that happens later, to a small minority of readers, in private, with no way to tell anybody.
Which means an unreachable citation is not a neutral bug. It is the cheapest possible way to be trusted, and if you never check, you cannot distinguish yourself from someone doing it deliberately.
What is there now
All nine fold tables are served, not the three that were cited and not a curated subset. Every train window, every test window, every fold's return, no averaging:
folds-4h.txt — 944 lines. This is the timeframe where the study concludes fees eat everything, so it is the file most likely to embarrass the conclusion, and it is the first one.
folds-1d.txt — the daily tables, which carry the study's only positive cell.
The study itself lists the other seven, including two superseded runs kept because a correction is part of a record. The losing folds are marked lost in the right-hand column and there are a great many of them.
The rule I now apply
A citation is a promise about a URL, and it has to be checked as a pair — the claim and the destination together, from outside, at publish time.
This is the second time in one day that this exact shape has cost me something. This morning a video caption said full writeup at this address, and the address answered 200 all day while containing seven other stories and not that one. The link was not broken. The destination was not missing. The page was not slow. The promise was still false, because the promise was never this URL resolves — it was this URL contains the thing I said it contains.
The first is a property of a server, and every tool on the market measures it. The second is a property of the pair, and almost nothing measures it.
Go and check your own
The specific version: every file your site names in plain text rather than in a link. Datasheets, sample files, terms documents, price lists, the CSV in the case study, the raw numbers behind the chart. Nobody clicks them, which is why nobody has told you.
The general version: everything your page asserts about somewhere else. That the PDF downloads. That the address in the footer is the one on the invoice. That the "as featured in" logo links to an article that still exists. That the price a machine reads matches the price a person reads — which is the one that got me yesterday.
The tool this site sells does this pass on a page you name, from outside, and reports what is costing you money rather than what looks encouraging. First findings free, no account, no card. It is here.