A project called Lexifina published something today (October 28) that lawyers have wanted for decades and programmers take for granted: a git-style history of the US Code. Pick a date, get the statute as it stood on that date. Every phrase colored by the Act that wrote it. git blame for federal law.
The claim that matters is not the interface. It's the verification method. For each pair of official release points from 2017 onward, they took the public laws enacted in between, replayed the amending instructions on the older text in date order, and compared the result against the next official version. That is a testable claim with published numbers, so I checked the numbers I could check.
The scoreboard they published
Across nine titles of the Code:
- 365 changes replayed
- 361 reproduced the next release point exactly — 98.9%
- The other 4 are shipped with a label saying how far they could be confirmed
And the effective-date breakdown, which is the part lawyers actually lose cases over:
- 261 changes took effect on enactment, by default
- 72 took effect per an effective-date note
- 29 took effect on a date written into the amendment itself
- 1 depends on an agency determination that may never come, and is left out of the text rather than guessed
That last choice is the most professional sentence in the post. A history that guesses an effective date is worse than a history with a hole in it.
One number that doesn't add up
261 + 72 + 29 + 1 = 363. The total is 365. Two changes are unaccounted for in the buckets as written.
This is probably a presentation slip — two changes may carry two kinds of date, or two may have been counted in the replay total but not yet dated. I can't tell from the post, and it doesn't dent the 361/365 result. But it's exactly the kind of thing this project exists to catch in other people's texts, so it's worth catching in theirs: if your product is "the numbers reconcile," your announcement should reconcile too.
The worked example checks out
Their example is the definition of a small business debtor, 11 USC § 1182(1), whose debt limit has been on a legislative roller coaster:
- Created by the Small Business Reorganization Act of 2019 at roughly $2.75 million
- Raised to $7.5 million by the CARES Act in March 2020
- The increase lapsed on March 27, 2022
- Congress restored it on June 21, 2022 — retroactively, for cases commenced on or after March 27, 2020 and still pending, reaching back over the gap
- It expired again in June 2024
- After the triennial adjustment, it now stands at $3,424,000
I verified the endpoint against two official sources that don't cite the blog. The US Trustee Program's Subchapter V page states the same history — SBRA enacted August 23, 2019, CARES raised the limit to $7.5 million, the extension expired June 21, 2024 — and gives the current figure as $3,424,000. The Federal Register adjustment notice (90 FR 8941, February 4, 2025) shows the same move in its table: the § 101(51D) amount goes from $3,024,725 to $3,424,000. The blog's number is right.
The trap the example exposes is real: for a case filed in May 2022, during the lapse, the release point printed at the time shows the lower limit. Only a later, retroactive Act changes the answer. The project admits it does not yet model that reach-back — on the 2022 gap dates it shows the lower limit with no warning. Honest limitation, clearly stated. It's also the single most dangerous date range in the whole dataset to get silently wrong.
Why this matters beyond law
The post cites a Technical University of Munich study from this year on LLMs answering questions about German statutes. Two failure modes: models apply rules superseded after their training cutoff, and they prefer the newer provision even when an older one governs the facts. One model answered 24% of post-cutoff questions correctly on its own, and 87% when the relevant statutory text was retrieved for it. A companion result is the uncomfortable one: when a model was asked to apply amending laws to earlier versions itself, the output looked 93–99% similar to the official text, but only half of single amendments — and a fifth of amendment chains — matched exactly, and experts found legally significant errors in the near misses.
93–99% similar, legally wrong. That is the failure mode of generated text everywhere, wearing a robe. The fix the Munich authors point to, and the thing this project actually built, is the same: temporal validity as a hard constraint, and a history that is checkable — replayed and diffed against the official record, not generated and eyeballed.
Lexifina's replay got 98.9% of changes byte-exact against the government's own next version, and labeled the other four instead of smoothing them over. That ratio — and the labeling — is the whole product. Every team shipping an LLM over documents with effective dates should be building the same thing: not a model that remembers the law, but a ledger that can prove which law was in force, and when, and which Act wrote the sentence you're reading.
Sources: Lexifina, "Git for the US Code" (Oct 28, 2026); US Trustee Program, "Subchapter V" (justice.gov); Federal Register 90 FR 8941 (Feb 4, 2025), via govinfo.gov. Bucket arithmetic and percentages computed from the post's own figures.
Top comments (0)