DocketXDocketX / Legal AI your malpractice carrier can live with
Legal AI your malpractice carrier can live with
Law firms want the speed of frontier models and cannot accept a fabricated citation in a filing. DocketRouter is a verification layer around any model: it grounds the answer in real Texas and federal law, checks every citation before you see it, and keeps a per-request record of what was retrieved, what was checked, and what was removed.
What the citation gate does
- Existence: every reporter citation is checked against a library of 615,391 Texas appellate decisions and an 18.1 million entry citation table across 989 reporters. Not found means removed and disclosed, never passed.
- Attribution: the case name attached to a citation is checked against the name the Texas courts use for that citation; a contradicted name is treated as fabricated.
- Quotations: quotations of the Texas and Federal Rules of Civil and Appellate Procedure and Evidence are checked verbatim against the current rule text; a quotation of superseded or invented text is removed and disclosed.
- Re-prompt, then strip: when a model invents, it is told which citations failed and asked to answer again with verified authority only; whatever still cannot be verified is stripped. At most two rounds, always disclosed.
- Audit trail: each request carries a request id with the retrieved sources, the verification report and the removed citations, retrievable from your account.
Deployment options
- API: OpenAI-compatible endpoint; keep your model choice, add the legal layer. Per-key spending caps, per-app attribution, usage export.
- Private inference: the same grounding and gate routed to a model on hardware the firm controls, so client text never leaves the firm. Built for firms that need real-time, in-house answers.
- Verification for your own AI: through a standard MCP endpoint, your agents check each citation against 18,123,788 reporter citations, test whether a case supports the proposition it is cited for, and pull court rules and statutes word for word, then search the Texas library by the hit. It sits beside whatever case-law search your agents already use.
Measured, not promised
- 348 answers across three models, 28 adversarial trap categories plus real cases per model: 0 unverifiable citations survived the gate, 0 citations invented before the gate on the grounded path.
- Benchmarked on blinded real Texas appellate cases, with a source-opinion exclusion so a model never sees the opinion the question was drawn from.
- Latency: a clean grounded answer is one model call plus a sub-second verification pass; streaming returns the verified sources before the first token.
What an audit trail looks like in practice
Each request carries an identifier, and against it we keep what the library returned, every citation the model produced, the result of each check, and anything removed with the reason. If a partner asks why a memo says what it says, the answer is retrievable rather than reconstructed. If a client asks whether their matter text went anywhere, the record shows which model served it and under what data policy.
Where this sits in your workflow
It is an API, so it sits behind whatever your firm already uses: a drafting tool, an internal chat surface, or a script a knowledge-management lawyer wrote. The model choice stays yours, including a private model on hardware you control, and the verification behaves identically across all of them. Nothing about it requires the firm to standardise on one vendor model.
What it does not do, so nobody is surprised
It does not tell you whether a verified case is still good law, and it does not confirm that a case supports the proposition it is cited for beyond checking whether the retrieved passage does. It holds Texas appellate law, Texas statutes, Supreme Court opinions and the federal and Texas rules, not federal district or circuit opinions and not other states beyond what is listed. Anything outside that returns unverified, which means our library cannot check it, not that it is fake.
Questions
Which models can we use?
Any major model through one API, including DeepSeek, Gemini, Claude and open-weight models, or a private model on your own hardware. The grounding and citation gate are the same for all of them.
Do you train on our data?
No. Prompts and answers are not used to train any model, and private inference keeps text inside your environment entirely.
What is covered today?
Texas appellate decisions 1840 through June 2026, Texas statutes, Texas court and county rules, the federal rules, and the U.S. Supreme Court. Washington is being added.
How is it priced?
Pay as you go on the API with per-key caps; library hits for your own agents at one cent per query; private inference and firm plans by agreement.
Related
Keep reading
DocketX products are research and explanation tools, not legal advice and not a law firm. They show you the real rule or opinion behind a statement so you can read it yourself. Coverage and counts are stated as of the index date on each product page.