Method
What this is
Ask what a signed acquisition agreement says about employee stock, break-up fees, or employees' pay and benefits after the deal, and get the clause quoted from the contract with a link to its source text. This is not legal advice. The system can be wrong; every claim shows its source so you can check it.
The agreements
Two sources. The first is MAUD, a dataset in which lawyers labelled 152 merger agreements with 92 question types in 39231 label rows. 100 of those agreements have published text; 88 of them are in the live index after 12 copies of tech deals were dropped as duplicates.
The second is technology-company acquisitions filed on EDGAR: the merger agreement attached to a current report, signed on or after 2015-01-01, whose target's industry code is in 3570–3579, 3661–3679, 7370–7379. They are chosen by that rule, not by hand: 318 agreements. The live index holds 406 agreements in 88628 passages, cut on article and section boundaries.
Which numbers a person checked
Numbers marked human-labelled (MAUD) are scored against the lawyers' labels in MAUD.
Numbers marked machine-built come from model passes, not lawyers. For the tech deals, two models (claude-opus-5-5 and claude-sonnet-5-5) each located the governing clause and stated the answer; only items where both agreed were kept. The judge model (claude-sonnet-5-5) is one of those two models, so it judged answers against a key it helped build; it may favour answers like its own, and its agreement rates should be read with that in mind. The lexicon that maps lay words to contract words was machine-built by claude-opus-5-5.
To test whether the machine-built key can be trusted, the same two-pass procedure was run on MAUD's own questions, so those questions have two keys; the Results page shows how often they agree. That is evidence, not proof.
What is not new
Retrieval over MAUD is a published benchmark: Pipitone and Houir Alami, LegalBench-RAG, arXiv:2408.10343v1. This project does not claim the task. It claims a working, measured, deployed system on top of that benchmark, plus three question families its labels do not cover: employee equity awards, break-up fees, and employees' pay and benefits after the deal. The Results page compares against the published baselines.
How an answer is made
The question is matched to one agreement by the company it names, or the one you pick. Inside that agreement, keyword search and a local embedding model (BAAI/bge-small-en-v1.5) each rank passages and the two rankings are fused. Lay words are rewritten into the agreement's own vocabulary, and each passage is shown with the definitions it depends on.
The model (claude-haiku-4-5-20251001) must return claims, each with a quote copied from a passage. A claim whose quote does not occur word for word in that passage is dropped; if none survives, the answer is "not stated in this agreement". A question that names no agreement, or several, gets "which agreement?" and no model call. Search shows the same ranked passages with no model call.
What it costs
The desk runs under a hard monthly cap. Hosting costs 7.09 US dollars a month; the model budget is the rest, 2.91, with a daily ceiling of 0.29. Prices were checked on 2026-10-05: 1.0 US dollars per million input tokens and 5.0 per million output tokens. The live model answers with extended thinking, as the evaluation runs did, with up to 4096 tokens of thinking per answer; the evaluation runs used the command-line tool's own thinking default. When the budget is spent, the desk shows cached answers only and says so. Search never spends model budget.
Stated limits
The tech half is public-company acquisitions: small private exits rarely file their agreements.
The tech-deal labels are machine-built. Two-pass agreement and the comparison with MAUD's lawyers are evidence, not proof.
MAUD's agreements are older than many of the tech deals, follow one annotation scheme and are not tech-specific, so results on MAUD may not carry over to the tech deals; nothing here measures that directly.
Lay questions are ambiguous. "What happens to my options" depends on vesting and on the agreement's own categories; the answer quotes the categories and does not pick one for you.
The answer evaluations ran through the Claude command-line tool, not the API the live desk calls. A calibration sample of 40 questions run both ways agreed on the answer state in 0.9 of cases (machine-built: the two runs were compared with each other, not with lawyers' labels).
Answer accuracy on MAUD was measured on an index of MAUD agreements alone. On the live index, which also holds the tech deals, the live rung's recall@5 on MAUD's questions is 0.5921 (0.573 to 0.6114) (human-labelled (MAUD)).
Ladder timings were measured on the development machine; the live server's own timings are on the Results page.
This is not legal advice.
What is logged
The service logs one line per request: the endpoint, the outcome, the time taken and the tokens used. It does not log your address or your question. Answers are kept in a cache with the question that produced them, so a repeated question costs nothing. Rate limits count requests per address in memory only.
Data and licences
MAUD (the Merger Agreement Understanding Dataset) is by The Atticus Project, under the Creative Commons Attribution licence; the dataset.
Agreements from EDGAR: the SEC states that its content "is considered public information and may be copied or further distributed by users of the web site without the SEC's permission". This site quotes clauses and links to each filing; the source repository ships the fetch script and filing identifiers, not the documents.
Source code: github.com/MichaelFornal/deal-terms-desk.