Words To Data

Your agent's favorite legal database.

words-to-data holds the U.S. Code, release point by release point, and the public laws that amend it. It records which amendment made which change, and the evidence for each record. An agent asks it questions with one command, words_to_data, and every answer comes with something it can quote.

Try it with SLEUTH

SLEUTH is the official agent. It runs in the pi coding agent. You need 6 GB of free disk space, and no Rust.

1. Install pi. The installer also installs Node.js when it is necessary.

   curl -fsSL https://pi.dev/install.sh | sh
   

2. Install SLEUTH into pi, start pi, and type /login to connect a model with a subscription or an API key. /sleuth-setup downloads the words_to_data command for your computer (Linux, macOS or Windows) and the example dataset (280 MB, 5.3 GB on disk), and opens the dataset. Then ask a question:

   pi install git:github.com/wordstodata/sleuth-pi
   pi
   /login
   /sleuth-setup
   Where in the Internal Revenue Code is the new Trump account section, and in which printing of the Code does it first appear?
   

The session is read-only. SLEUTH finds the section from the words alone, and it shows the evidence. A real answer (2026-10-23), cut:

Where: the new Trump account section is 26 U.S.C. § 530A ("Trump accounts"), in a brand-new part IX ("TRUMP ACCOUNTS") added at the end of subchapter F of chapter 1: uscode/title_26/subtitle_A/chapter_1/subchapter_F/part_IX/section_530A

First printing: the 2025-07-30 printing of Title 26. w2d_path reports the section present at 2025-07-30, 2025-08-14, 2026-01-23 and 2026-07-12, … a search for "Trump account" at 2025-07-18 returns 0 fields.

Source: OLRC classifies part_IX/section_530A as new from Pub. L. 119-21 § 70204(a)(1) (bill 119-hr-1) …

Then ask more questions, for example:

• What does 26 U.S.C. § 280C(c)(1) say now, and what did it say before the One Big Beautiful Bill Act changed it?
• In Boeing Co. v. United States, what did Justice Thomas say at 537 U.S. 459, and which Code sections does his dissent cite?
• BMW (US) Holding Corp. v. United States (Court of Federal Claims, August 2026) cites the bonus depreciation rule, 26 U.S.C. § 168(k). Has § 168(k) changed since the decision?

The example dataset holds:

• The U.S. Code, every title, at 5 printings: 2025-07-18, 2025-07-30, 2025-08-14, 2026-01-23 and 2026-07-12.
• 7 bills of the 119th Congress, with their amendments linked to the changes in the Code, and the House roll calls. Two of them are H.R. 1 (Pub. L. 119-21, the One Big Beautiful Bill Act) and S. 1071 (the NDAA for FY2026, Pub. L. 119-60).
• 49 court writings in 28 cases, with pages, paragraphs, footnotes and their citations of the Code.

Why agents like it

• Evidence, not recall. A link says what made it: the address in the bill's markup, the window it changed in, and the words the law quotes. settle --explain prints all of it. The agent quotes the record and does not rely on what it remembers about a law.
• No model inside. The dataset is built by rules from the publisher's markup and the OLRC's classification tables. What the rules cannot place goes on a work list (residue) with the reason it stopped. An answer is never a guess that looks like data.
• It says what it does not know. cases-citing names each period the dataset holds no printing for, and says that this is not a statement that nothing changed. residue gives the reason each unlinked amendment stopped.
• Output that fits in context. Listing commands stop at 20 rows and say how many they did not show, and how to get the rest. --path, --work, --at and --bill narrow a listing, and --json gives machine-readable rows.
• Writes that are hard to get wrong. link-amendment refuses a path that did not change. settle records a review as its own record, and nothing is deleted. A wrong link stays visible, and the newest review stands.

Give it to your agent

SLEUTH, the official agent

SLEUTH-pi is a package for the pi coding agent. It has a tool for each words_to_data command, the playbook as four skills (answer, residue, review, build), and a write gate. A session is read-only unless you give it --write. Try it with SLEUTH gives the steps to install it.

pi install git:github.com/wordstodata/sleuth-pi
pi
/dataset law.sqlite

Any other agent

Give the agent three things: the words_to_data binary, a dataset, and docs/agents/working-a-dataset.md. The playbook tells it how to answer a question with evidence, how to work the residue, and how to review links. It needs no other tools.

In the test of the playbook (#267), six new agents answered four questions about Public Law 119-21 correctly, in 17–22 tool calls each, and worked its residue from 55 items to 3. Every answer was checked by hand.

Build a dataset

You need about 3.5 GB of disk and a free Congress API key. Each release has the command for Linux (x86_64, aarch64), macOS (Apple silicon, Intel) and Windows (x86_64). In the file name below, use aarch64-apple-darwin (Apple silicon) or x86_64-apple-darwin (Intel) on macOS. With Rust, cargo install words-to-data also works.

curl -fL https://github.com/wordstodata/words-to-data/releases/latest/download/words_to_data-x86_64-unknown-linux-musl.tar.gz | tar xz
sudo mv words_to_data /usr/local/bin/
export CONGRESS_API_KEY=<your key>

words_to_data build-dataset law.sqlite --uslm-dates 2025-07-18,2025-07-30 --bills 119-hr-1
words_to_data add-classifications law.sqlite
words_to_data redesignations law.sqlite --bill 119-hr-1 --between 2025-07-18 2025-07-30
words_to_data link-by-evidence law.sqlite

That is the whole Code at two release points, and Public Law 119-21 (the reconciliation act of July 2025) linked to the changes it made. It takes about a minute and a half.

What your agent sees

Three questions, and the real output. Where is section 174A, which amendment made it, and why should anyone believe that?

$ words_to_data search law.sqlite "domestic research or experimental" --work uscode/title_26 --at 2025-07-30
...
uscode/title_26@2025-07-30  …/subchapter_B/part_VI/section_174A  [heading]  1 occurrence
    Domestic research or experimental expenditures

$ words_to_data annotations law.sqlite --bill 119-hr-1 --path uscode/title_26/subtitle_A/chapter_1/subchapter_B/part_VI/section_174A
  [Pending] uscode/title_26 2025-07-18 -> 2025-07-30  insert 119-hr-1 amd 2841e731435d …
      (a) In General.-Part VI of subchapter B of chapter 1 is amended by inserting after section 174 the f…
      link 7cdb05c5323a  …/section_174A  [quoted_words]

$ words_to_data settle law.sqlite --link 7cdb05c5323a --explain
How it was made:
  Source: rule:evidence_matching
  Method: address, window and quoted words@9
  Address: /us/usc/t26/s174A, from the bill's markup
  Window: uscode/title_26@2025-07-18 to 2025-07-30
    the first window after the law's enactment on 2025-07-04 in which something under the
    address changed.
  Changes under the address: 1
  Chosen: by quoted words
    the change shows the words the bill quotes: enacted text 1 of 1.
...

words_to_data residue law.sqlite --bill 119-hr-1 lists what the rules could not link, and why. words_to_data --help lists every command.

Use it as a library

[dependencies]
words-to-data = { version = "0.7", default-features = false }

use words_to_data::dataset::Dataset;

let dataset = Dataset::open_sqlite("law.sqlite")?;
let path = "uscode/title_26/subtitle_A/chapter_1/subchapter_B/part_VI/section_174A";
for annotation in dataset.annotations_for_path(path)? {
    println!("{:?} by {}", annotation.operation, annotation.source_bill.bill_id);
}
// Insert by 119-hr-1

default-features = false leaves out the command line and its dependencies.

More

• Grow a dataset in place, with no rebuild: add-release-points and add-bills. A grown dataset holds the same links as one built from scratch. release-points lists the release-point dates the mirror offers and the cache holds.
• Court opinions that cite a provision: add-opinions (needs COURTLISTENER_API_KEY), then cases-citing.
• Everything is cached, under your user cache directory. --offline reads only the cache.
• Concepts: CONTEXT.md is the glossary. docs/adr/ records each design decision and its reason. Start with 0013, how a change is linked.

Development

cd tests/test_data && tar xf test_files.tar.xz && cd ../..   # the test corpus is an archive
cargo test

Without the archive, about 190 tests fail on missing files. Tests use real committed data only. CLAUDE.md holds the working rules, including the merge queue.

License

MIT or Apache-2.0, at your option.