The situation
Asked a regulatory question, an AI model answers from what it learned in training, and the paragraph or the wording can be off. With web search it gets there, but by reading far more.
What Joakim did
- Built MetLex, which gives an AI tool the verbatim text of DORA, its technical standards, the GDPR and the FFFS, with citation, official link and snapshot date.
- Made it a service, so Mettle can keep the texts current and add to them without access to the systems of those who use it.
- Tested it in October 2026 with one model and 39 regulatory questions, each asked once from memory, once with web search and once with MetLex.
- Graded against the official text as built into MetLex, and stated the limits: Joakim wrote the questions, and Mettle sells MetLex.
What came of it
With MetLex the model found the right provision as often as with web search, in about a third of the time (15 s against 48 s, median), and linked the official source every time.