The DocketRouter legal model
Every request through DocketRouter passes through our legal model before and after the enterprise model answers. Before: it retrieves the on-point Texas and federal rules, statutes and opinions from our fused legal index and briefs the model on exactly what was found. After: it checks every citation in the answer, tells the model the truth about each one and asks it to re-answer, then removes what still cannot be verified and says so. Citations that cannot be checked are reported as unverified rather than presented as confirmed, and if the index is unreachable the response says so instead of answering as though it were grounded.
Teal = the DocketRouter layer (our corpus, index and verifier). The enterprise model does the reasoning; we supply verified law, tell the model which of its citations failed, and remove what still cannot be verified. Dotted = feedback: re-answer on a failed check, and measured results tune retrieval.
The truth loop
Four stages, run on every grounded request.
- 1. Find the law
The question, not the whole record, is the query. Sending a full record drowns the legal signal in party names and dates, so the index sees the question and the model sees the facts. One fused index on our own hardware answers in about a second: 615,391 Texas decisions, 122,681 Texas statute sections, 45,660 U.S. Supreme Court opinion documents, 370 Texas county and local rules, 866 Texas court rules, 227,809 California case law, 101,200 New Jersey case law, 49,830 Virginia case law, 87,464 Washington case law, 63,362 Minnesota case law, 14,914 Alaska case law, 23,434 Delaware case law, 35,818 Idaho case law, 57,743 Kansas case law, 31,116 Maine case law, 101,361 North Carolina case law, 23,554 North Dakota case law, 26,773 Rhode Island case law, 25,591 South Dakota case law, 34,885 Utah case law and 55,991 Wisconsin case law, alongside 1,413 Texas rules of court and 595 federal rules held verbatim and quoted by short cite.
- 2. The model answers
The enterprise model you chose reads the retrieved authority and writes the answer. It is briefed on exactly what was found, told never to invent a case, citation, quotation or holding, and told to say when something could not be verified.
- 3. Check every citation
Every citation in the answer is checked, up to 80 per answer, our index first and the public-domain citation table second. If any fail, the same model is shown the full picture, which citations were confirmed and which could not be verified anywhere (a miss is reported as unverified, never as nonexistent), and is asked to rewrite in full, up to two rounds. A citation that appears in your own record is never treated as fabricated.
- 4. Remove what fails, and say so
Whatever is still unverifiable after the last round is removed mechanically, the surrounding statement is left in place and marked unsupported, and a note at the end of the answer says how many citations were removed. The removed list is in the verification report. Nothing unverified ships as if verified.
The in-house model, and where it stops
- Runs on our own hardware, next to the index.
- Grounds the input: ranks and expands the retrieved authority so the candidates the enterprise model sees are the authorities that opinions on that issue actually rely on.
- Checks the output: every citation, against our index first.
- Under a private-pod key, it is the only model your prompt ever reaches.
- It never writes the customer-facing answer. The enterprise model re-answers; the in-house model grounds and checks.
- It never rewrites the model's prose. Removal is mechanical and limited to the citation span, and the answer says what was removed.
- It never alters a client document. Evidence is fenced verbatim; only hidden characters that carry no legal meaning are stripped.
- Federal Rules of Civil Procedure: 395 chunks, verbatim
- Federal Rules of Evidence: 95 chunks, verbatim
- Federal Rules of Appellate Procedure: 105 chunks, verbatim
- Texas Rules of Civil Procedure: 937 chunks, verbatim
- Texas Rules of Evidence: 122 chunks, verbatim
- Texas Rules of Appellate Procedure: 354 chunks, verbatim
- Fused legal index: 615,391 Texas decisions, 122,681 Texas statute sections, 45,660 U.S. Supreme Court opinion documents, 370 Texas county and local rules, 866 Texas court rules, 227,809 California case law, 101,200 New Jersey case law, 49,830 Virginia case law, 87,464 Washington case law, 63,362 Minnesota case law, 14,914 Alaska case law, 23,434 Delaware case law, 35,818 Idaho case law, 57,743 Kansas case law, 31,116 Maine case law, 101,361 North Carolina case law, 23,554 North Dakota case law, 26,773 Rhode Island case law, 25,591 South Dakota case law, 34,885 Utah case law and 55,991 Wisconsin case law
- The public-domain citation table: citation lookup, after our index
- The served index holds 615,391 Texas opinion documents: the combined and lead opinions that carry the court's holding. The open dataset docketx/us-caselaw-tx publishes 662,717, which additionally includes 38,008 separately filed concurrences, dissents, rehearings and addenda, and 9,318 very short documents below the served build's 500-character floor. Both numbers are measured; they count different things.
- Verbatim rule text, cited by short cite
- Related opinions with dates and links
- Citation verdicts: found, not found, ambiguous, unverified
- Client files fenced as evidence, never as instructions
- A plain statement when retrieval was degraded, so it does not answer as if grounded
- Sources: the rules and opinions that were retrieved
- Verification report: every citation checked, and the list of any removed
- Injection report: clean, suspicious or hostile per client document
- Degraded flag when the index was unreachable or nothing matched
- Exact cost, with the upstream cost beside it in your usage export
Your data
Legal records are routinely written by the other side. The layer is built on that assumption.
Uploaded documents are treated as evidence, never as instructions: fenced, hidden-character smuggling stripped, and scanned for embedded instructions. The scanner is tuned against real legal prose so ordinary filings do not trip it. The verdict for each document is in the response.
A key can be set to a policy where only in-house models are allowed. Under that policy the prompt never leaves the box, and routing to providers that retain or train on data is denied.
Each key carries a hard spend cap that is enforced upstream. Spend can never exceed the cap.
A matter holds a message history and its attached case files, encrypted per client, and narrows retrieval to that matter's documents.