The Web Address That Carried Its Own Prefix Twice

Fifty-five of our research write-ups are each built on one published paper, and every paper has a permanent web address called a DOI. We print that address on the page so you can go and read the paper yourself.

We ask whoever writes the record to store the address in its short form. The page then adds the resolver in front of it. On nine of our research pages the address had been stored in its long form already, so the page added the resolver to an address that had one, and printed the prefix twice.

The link still worked. A reader who clicked it landed on the correct paper, which is why nobody caught this for months. What was wrong was what you could see, and what a machine was handed: the same doubled address was published in our structured data as the paper’s permanent identifier, and a permanent identifier that does not match the real one is not much of an identifier.

The page now trims the address to its canonical form before it prints it or hands it to a machine, whatever shape the record stores it in. The nine records still hold the long form and are being corrected separately.

The second thing we found is smaller to look at and closer to the bone. We publish a document that describes our own data to AI systems, field by field, so a model reading us does not have to guess. One line of it described the summary we write for each chart as the caption you see underneath that chart. It is not. It is the description we hand to structured data, and the caption you read is written separately inside the chart itself. On most of our charts the two say different things.

Nobody was misled by a caption. What was wrong was the contract, and a contract we publish for machines is worth exactly what its accuracy is worth. The line now says what the field actually does, and points at the field that actually holds the caption.

Both of these were caught by us, reading our own pages, while rebuilding something else.