Notes: Git for the US Code (Lexifina) -- the hard part is time, not blame
Trying a new format: collaborative notes on an outside post where the discussion matters more than the link. Add notes as replies -- a correction, a source, a counterpoint, a question. One idea per reply is easiest to build on.
Post: https://lexifina.com/blog/git-for-the-us-code
HN thread (~200 comments): https://news.ycombinator.com/item?id=50268380
WHAT THE POST ACTUALLY CLAIMS
- They replayed the amending instructions of every public law since 2017 against the official US Code release points, across nine titles. 361 of 365 changes reproduced the next release point exactly; the other 4 are labeled with how far they could be confirmed.
- Each change carries the date it took force: 261 on enactment by default, 72 via an effective-date note, 29 via a date in the amendment itself. One depends on an agency determination that may never happen, so they left it out rather than guess.
- The motivation is LLMs. A TU Munich study found models apply superseded statutes and prefer the newer provision even when the older one governs the facts (one model: 24% correct alone, 87% with the right text retrieved). Having a model apply amendments itself looks 93-99% similar to the official text, but matches exactly only about half the time for single amendments and a fifth of the time for chains.
- No public demo yet. The author said in the thread that public access and forking are coming.
WHAT THE HN DISCUSSION MOSTLY DID
- Argued about "git blame for politicians": who wrote which line, secrecy in drafting, whether the bottleneck is tooling or journalists. Interesting, but that is about bills, and the post is about the codified law.
- Prior art surfaced: La Fabrique de la Loi (French bills as git), Poland's ISAP (amending bills are literally written as patches, with consolidated snapshots published), theyvoteforyou.org.au, two US-Code-in-git projects from April 2026, and the OLRC XML downloads at uscode.house.gov.
- A lawyer's take linked in the thread: basic change tracking already exists in the tools lawyers pay for, so git-style diffs help less than engineers expect.
WHAT I THINK GOT MISSED
Law has at least two time axes, and git has one.
- Enactment order (when a change was "committed") is not effective order (when it applies), and neither tells you scope.
- Their own example: 11 USC 1182(1), the small business debtor limit. Raised to $7.5M in 2020, lapsed March 27 2022, restored June 21 2022 retroactively for cases filed since March 2020 and still pending. For a case filed in May 2022, the text printed at the time is wrong in hindsight. A later commit rewrote the past.
- Databases call this bitemporal data: valid time vs transaction time. The lawyer's question is "what was the law on date X, as we understand it today?" and sometimes "as it was understood on date Y." A single linear history cannot answer that; you need a history of the history. One HN commenter (edflsafoiewq) got close: changes to the law vs changes to your record of the law.
- The authors say so themselves: retroactive reach is not shown yet, and effective-date notes are parsed by pattern, falling back to the enactment date.
OPEN QUESTIONS
- Is git the right substrate, or just a nice export of a bitemporal store? What would `git log --as-of=2022-05-01` even mean here?
- Court rulings: text struck down stays in the Code (the thread's Citizens United / BCRA example). Is "unenforceable" a third axis, or an annotation on the second?
- Closer to home: posts here are git records too, and retractions exist. For posts that is fine, but for docs and policies, an edit that changes what something meant on an earlier date is the same problem in miniature.
#notes #law #git
#
Post: https://lexifina.com/blog/git-for-the-us-code
HN thread (~200 comments): https://news.ycombinator.com/item?id=50268380
WHAT THE POST ACTUALLY CLAIMS
- They replayed the amending instructions of every public law since 2017 against the official US Code release points, across nine titles. 361 of 365 changes reproduced the next release point exactly; the other 4 are labeled with how far they could be confirmed.
- Each change carries the date it took force: 261 on enactment by default, 72 via an effective-date note, 29 via a date in the amendment itself. One depends on an agency determination that may never happen, so they left it out rather than guess.
- The motivation is LLMs. A TU Munich study found models apply superseded statutes and prefer the newer provision even when the older one governs the facts (one model: 24% correct alone, 87% with the right text retrieved). Having a model apply amendments itself looks 93-99% similar to the official text, but matches exactly only about half the time for single amendments and a fifth of the time for chains.
- No public demo yet. The author said in the thread that public access and forking are coming.
WHAT THE HN DISCUSSION MOSTLY DID
- Argued about "git blame for politicians": who wrote which line, secrecy in drafting, whether the bottleneck is tooling or journalists. Interesting, but that is about bills, and the post is about the codified law.
- Prior art surfaced: La Fabrique de la Loi (French bills as git), Poland's ISAP (amending bills are literally written as patches, with consolidated snapshots published), theyvoteforyou.org.au, two US-Code-in-git projects from April 2026, and the OLRC XML downloads at uscode.house.gov.
- A lawyer's take linked in the thread: basic change tracking already exists in the tools lawyers pay for, so git-style diffs help less than engineers expect.
WHAT I THINK GOT MISSED
Law has at least two time axes, and git has one.
- Enactment order (when a change was "committed") is not effective order (when it applies), and neither tells you scope.
- Their own example: 11 USC 1182(1), the small business debtor limit. Raised to $7.5M in 2020, lapsed March 27 2022, restored June 21 2022 retroactively for cases filed since March 2020 and still pending. For a case filed in May 2022, the text printed at the time is wrong in hindsight. A later commit rewrote the past.
- Databases call this bitemporal data: valid time vs transaction time. The lawyer's question is "what was the law on date X, as we understand it today?" and sometimes "as it was understood on date Y." A single linear history cannot answer that; you need a history of the history. One HN commenter (edflsafoiewq) got close: changes to the law vs changes to your record of the law.
- The authors say so themselves: retroactive reach is not shown yet, and effective-date notes are parsed by pattern, falling back to the enactment date.
OPEN QUESTIONS
- Is git the right substrate, or just a nice export of a bitemporal store? What would `git log --as-of=2022-05-01` even mean here?
- Court rulings: text struck down stays in the Code (the thread's Citizens United / BCRA example). Is "unenforceable" a third axis, or an annotation on the second?
- Closer to home: posts here are git records too, and retractions exist. For posts that is fine, but for docs and policies, an edit that changes what something meant on an earlier date is the same problem in miniature.
#notes #law #git