alea-institute/alea-draft-checklist

draft_checklist: practical drafting from a real public document

Each document pairs a real public-record source with a drafting task and the finished draft. The sources are a court opinion or order, a Federal Register document, an agency web page, a congressional committee report or hearing excerpt, an SEC filing, or a contract. The task names a reader and a kind of text and gives a word limit. Kinds of text: a client letter on what the court decided, an internal memo, a notice to employees or members, an FAQ entry, a plain-English summary of an agreement's key terms, a confirming e-mail, or a note to the board. In about 60% of documents the instruction lists the required items explicitly (a date, an amount, a party, an outcome). In the rest it says only "include the dates, amounts and parties the reader needs".

The set is for mid-training. It teaches a language model to draft short practical documents that:

• stay faithful to a source;
• keep every date and amount attached to the right thing;
• add nothing the source does not say: no outside knowledge, no invented next steps, no assumed client relationship.

How the data was made

1. Checklist. A program detected identifiers in the source (agency, court, case name, docket, comment and effective dates). A language model proposed the key facts as short exact spans. Each item was kept only if its value appears verbatim in the source.
2. Drafting. GLM-5.3-Flash wrote most drafts and gpt-6-luna wrote a minority. The drafting rules: use only the source; keep status, conditions and negations exact; write in your own words; stay well under the limit; use no outside knowledge, no placeholders and no invented next steps.
3. Verification. A draft is kept only if it passes every check below:
   • Program checks (drafts that fail are dropped, never edited):
     • every numeric checklist value is present;
     • no date or dollar amount appears that is not in the source;
     • the draft is under the word limit;
     • no 20-word passage is copied;
     • no contact details or URLs;
     • no court named that the source does not name;
     • no party or state labels from outside knowledge;
     • no "attorney for ..." relationship the source does not establish;
     • no placeholders or misspelled source names;
     • no "filed today"-type claims;
     • no first-person promises of next steps.
   • Two independent judges from a different model family than the writer:
     • DeepSeek-V4.1-Flash (or GLM-5.3-Flash for luna drafts) checks that every required item is covered and that nothing is unsupported;
     • a second, strict DeepSeek fact-check flags any characterization not literally supported by the source (omissions alone are allowed).
4. Decontamination and balance. Rows overlapping evaluation and held-out sets were removed (8-gram and 13-gram screens), and no source pool exceeds 30% of documents. 1.5% of source documents are held out and not released.

Counts and composition

• 307 documents, 0.32M tokens (alea-v2b-sw32k tokenizer), one source document each.
• Source pools: contracts 29.6%, agency web pages 28.8%, court orders 10.4%, court opinions 9.7%, 10-K excerpts 7.4%, congressional hearings 6.2%, committee reports 4.5%, 8-K reports 2.9%. No single court, agency, website or filer exceeds 3% of documents.
• Kinds: memo 25%, e-mail 17%, FAQ 16%, term summary 14%, notice 13%, note to the board / staff 10%, client letter 6%.

Review

The dataset authors read 30 rows against their sources. A separate blind reviewer then checked three rounds of 60 rows, the last two on fresh rows:

• Round 1: 4 errors.
• Round 2: 10 errors (mostly outside knowledge, mislabelled numbers, invented instructions or relationships).
• Round 3: 1 error.

After each round the generator and the checks were tightened (the second strict judge and the outside-knowledge checks were added), and every row with an error was removed. Round 3 also noted generic sign-offs and some heavy paraphrase of the source.

Limitations

• The set is small: strict checking keeps only about one draft in six.
• Some drafts stay close to the source wording, and sign-offs are often generic.
• English only.