Lexifina rebuilds the US Code as a dated edit history for LLMs

A legal workspace called Lexifina has replayed every public law amendment on nine titles of the US Code. The result is a dated edit history. Alan Yahya published the method on the Lexifina blog on 28 October 2026. The post reached 320 points on Hacker News by mid-day.

The problem is simple. A lawyer wants to read a statute as it stood on a given date. The US Code is published as a series of release points. Each one is current through a named public law. Between two releases, the only record of what changed is a set of amendment notes written in prose. A model trained before the latest release does not know the newer rules at all.

What they built

The team took each pair of release points from 2017 onward. For every pair, they took the public laws enacted in between. They carried out those amending instructions on the older text in date order. Then they compared the result with the newer release.

Of 365 changes across the nine titles, 361 reproduced the next release point exactly. The other four are shown with a label saying how far they could be confirmed.

Each change carries the law that wrote it and the date it took effect. The post says 261 took effect on enactment by default. 72 took effect by a note in the law. 29 took effect by a date in the amendment itself. One depends on an agency determination that may never come. Lexifina leaves it out of the text rather than guess.

Why models get the dates wrong

The post cites two academic papers. A Technical University of Munich group found two failures in models answering questions on German statutes. First, they apply rules that were superseded after their training cutoff. Second, they prefer the newer rule even when an older one governs the facts.

One model answered 24 percent of post-cutoff questions correctly on its own. The same model answered 87 percent when the right statutory text was retrieved for it. The authors conclude that temporal validity has to be treated as a hard constraint.

The same Munich group tried asking a model to apply amending laws to earlier versions. The output looked 93 to 99 percent similar to the official text. But only half of single amendments matched exactly, and only a fifth of chains of amendments did. Experts found legally significant errors in the near misses.

What a reader can now do

A user picks a date and gets the text in force then. Words that read differently today are marked. Each phrase is coloured by the Act that wrote it. A change to a defined term lists every other provision that uses it.

The post states its own limits plainly. It does not yet show a later law's reach back on earlier dates it covers. Effective-date notes are read by pattern, so unusual wording falls back to the date of enactment.

A team building a legal assistant with a language model needs the data layer in this shape. The text store has to carry a date on every passage. Retrieval has to pin the answer to the date in the question. The model must refuse to answer when the dated text is missing. A plain web crawl of the US Code gives none of that.

Source

• Alan Yahya, Git for the US Code, Lexifina blog.
• Hacker News discussion, 320 points.