tercio14 is my project for people preparing the Spanish Guardia Civil entrance exam on their own. A good part of what it does is unglamorous: it follows the official notice board, keeps a dated chronology of each year's process, and links the documents — the call, the lists, the resolutions — so that you can read the source instead of somebody's summary of it.

That only works for as long as the source is there.

Two out of thirty-four

At the end of September I closed out the record for this year's process, and while I was in there I went back to the archived boards for 2024 and 2025 to check their links. Of the thirty-four documents the 2024 board points at, two still answered from the official server. Two. The rest had moved, or been renamed, or been swept away when the department rebuilt the page they lived under.

Nothing about those documents stopped mattering when they stopped resolving. A process that ran two years ago is the best available evidence of how the next one will run: what the medical review actually checked, how long each stage took, what the wording of a resolution looked like. Somebody reading my chronology of 2024 was reading a carefully dated history in which most of the footnotes were 404s.

What the archived boards look like now

Every document on the archived boards for 2024, 2025 and this year now carries a "copy on archive.org" link behind the original. The original stays first: if the institution is still serving the document, that is where you should read it.

How it worked out:

Year Documents With a copy
2024 34 15
2025 37 32
2026 35 35

This year is a clean sweep because I captured those while they were still being served, which is the whole point and also the only part of this I control. The twenty-four without a copy aren't a gap in the process: the origin doesn't serve them and the Internet Archive has no capture of them under any form of the URL I could construct. They keep their original link, which is honest if not useful.

Two things I got wrong on the first pass

I asked the Internet Archive the easy question first — "do you have this?" — and believed the answer. It said no for documents it demonstrably had. Asking its index for the most recent capture that returned a 200, rather than asking whether a copy was available, turned up copies for a good number of the documents the easy question had written off.

The second mistake was assuming a document has one URL. The same PDF gets published under a couple of different host and path shapes, and the one printed on the board isn't necessarily the one a crawler ever saw. Trying the handful of known shapes recovered eight more copies from 2024 alone.

There is also a detail that changes nothing for a reader and that I would defend anyway: I made the archived copy of a board come out byte-identical regardless of which of two routes produced it. That file is the reference copy of the year — the thing that is left when the institution rebuilds its site — and its contents shouldn't depend on which source happened to be reachable the afternoon I generated it. That is the class of drift you don't notice until it's the only copy you have.

The uncomfortable part

If what your product offers is that it points at somebody else's documents, then those documents disappearing is your problem, not theirs. The institution isn't doing anything wrong by reorganising its website. It never promised me a permanent URL, and nobody there knows I exist. The permanence was something I assumed because the links worked on the day I wrote them down.

What bothers me more is how I found out. Nothing in my setup would have told me: there is no alarm for "a document you linked two years ago is gone", and a page full of dead links renders exactly as well as a page full of live ones. I found out because closing the year gave me a reason to re-read links I would otherwise never have revisited. Two of thirty-four had been dead for who knows how long, on a page that looked completely fine.