In a country like Finland, what the law says is a function of applying all the amendment acts over a hundred+ years of history and the consolidated "current law" doesn't have any legal force. One can just do this manually and consolidate to the current state, but it's unreliable. Same reason that logarithm books were hard to compile without errors. So it's extremely difficult to know what the law SAYS with certainty. Forget semantics, even what the second part of first section of Foo Law says. This is what LawVM solves. No LLMs or AI involved (except for corrigendum matching in the Finland pipeline - literally an AI-hard problem to read visually from PDFs - can substitute in humans too for this task though)
For other jurisdictions, the shape of the problem is different, LawVM may only verify consistency. Or be used to help determine bitemporal point-in-time queries with exactness. What did XYZ law say at ABC time?
LawVM functions kind of like LLVM - there are different frontends for different jurisdictions, and all use the same core and IR. While UK, Estonia, Sweden and Norway frontends also exist in various states, the most advanced frontend is Finland. It replays Finnish amendment acts from the official Finlex Statute Collection, compares against the ~unofficial/without-legal-force provided-as-is Finlex consolidated text, and classifies divergences. So far I've reported 22 high-confidence candidate findings to Finlex and LawVM surfaces hundreds more. It turns out that absent discipline that a system like what LawVM requires, everything drifts. Sometimes the XML is inconsistent with the published PDF version of the law and many other such disagreements. Law SHOULD be knowable, always. And yet it's not, at least not always.
The project is v0.1: useful as a construction proof and audit substrate, not legal advice or an official consolidation.
It is quite an ordeal to parse natural language, but it is possible. Even a difficult-to-compile language is still possible to compile. Helps that legal text is quite regular compared to freeform natural language.
With something like this, it becomes possible to approximate certainty in "what the law actually says" (or consistency, in case of Estonia and others where the consolidated law HAS legal force directly).
I'm hoping gets adopted by legal publishers, public institutions, researchers, or civic infrastructure groups. Version control became normal for software; legal text-state needs comparable level of discipline.