How it works
Three answers to the same question, on one screen.
A fluent answer and a correct answer look identical. This is built so you can tell them apart, including the parts that are not flattering.
The problem
Fluent is not the same as correct.
An AI answer about immigration law reads the same whether it is right or wrong. The tone is even, the citations look like citations, and nothing in the writing tells you which sentences are safe to rely on.
It is a measured problem. Researchers at Stanford's RegLab, publishing in the Journal of Empirical Legal Studies, found tools built specifically for legal research producing hallucinated content on roughly one query in six. General-purpose chatbots did considerably worse.
Asking a second AI helps less than it sounds, because leading models are not independent of each other. On open questions, two frontier models have been measured landing on the same wrong answer about one time in eight. So agreement here is treated as supporting evidence and never as proof, and the screen says so every time. The value of a second opinion is not the reassurance when it matches. It is the warning when it does not.
The design
Every question produces three things at once.
You cannot receive one without the others. The pattern is borrowed from safety-critical engineering: run two channels that fail in different ways, then compare them. Here the comparison happens in your field of vision rather than quietly behind the screen.
Panel A
Sourced answer
Works over the curated library and, where the question warrants it, live official sources. A citation on every claim. A mechanical check then matches the quotations and government links against what was actually retrieved.
Panel B
Independent analysis
Answers from training knowledge alone. No library, no live lookup, no sight of Panel A. Labelled as such, and stamped with the knowledge cutoff of the engine behind it.
Panel C
Where they agree, and where they do not
Reads both and reports the convergences and the divergences. Instructed never to call agreement confirmed or verified, because two models can be confidently wrong together.
Panel B is not a footnote to Panel A. Its independence is enforced in code: it never receives Panel A's answer or its passages, and each panel remembers only its own prior answers. Co-equal in the layout does not mean interchangeable in weight. The sourced panel is the one designed to be relied on. The independent panel is there to disagree with it.
One variable
Set both panels to the same engine.
Do that and the two panels run an identical model on an identical question. Every difference of style, training and temperament cancels out. One thing remains: Panel A searched the library and had its quotations and links checked, and Panel B did nothing of the kind.
Whatever happens next is attributable to retrieval and checking, and to nothing else. Several providers are integrated, so when one fails the panel reports the outage rather than quietly answering from somewhere else.
The library
Where the law comes from.
Not the open web. A curated library of primary U.S. immigration law, 1,479 documents in 55,495 passages, indexing 7,000+ distinct authorities, plus live official sources at the moment you ask when the question calls for it. Every passage carries an authority label, a source tier and a quotation rule, with no gaps.
| Source | Documents | Passages | How it reaches the library |
|---|---|---|---|
| USCIS Policy Manual | 456 | 14,637 | Structured ingestion, no public API published |
| Board of Immigration Appeals precedent | 448 | 14,077 | CourtListener, approved commercial licence |
| INA, as codified at 8 U.S.C. | 210 | 4,184 | govinfo, from the Government Publishing Office |
| 9 FAM, the visa volume | 149 | 4,820 | Structured ingestion, no public API published |
| 8 CFR | 132 | 6,057 | eCFR REST API |
| Federal Register | 44 | 1,482 | Official REST API, polled every morning |
| The Dickie Woods database | 28 | 9,500 | Curated and maintained in house |
| 22 CFR, the visa parts | 9 | 639 | eCFR REST API |
| Attorney General precedent | 2 | 22 | CourtListener, approved commercial licence |
| DOS Visa Bulletin | 1 | 77 | Official snapshot; priority dates always confirmed live |
| Total | 1,479 | 55,495 | Counted in production on 23 August 2026 |
Every question is searched two ways at once, by meaning and by exact wording, and the two rankings are fused, so ordinary language and a quoted term of art both find the right passage. Fast-changing material, priority dates, processing times and fees, is flagged and never trusted from the library: it forces a live lookup instead. Official feeds are polled every morning at 09:00 Singapore time, and a document is kept on structural grounds rather than a keyword guess.
Guardrails
What stops a made-up answer.
No single mechanism prevents an AI from inventing a case. Several are layered, each aimed at a different way an answer goes wrong. The fail-closed gate is the one that costs us money: when nothing relevant is found, the platform declines to answer under sourced framing, says so, and refunds the question.
| Layer | The failure it catches |
|---|---|
| A curated library of primary law | An answer assembled from whatever the open web said that morning |
| An allowlist of authorised official hosts | A live search wandering onto a forum or a marketing page and treating it as authority |
| A citation required on every claim | Unsupported assertions, and commentary weighed as though it stood level with the statute |
| The fail-closed gate | Nothing found, so the engine answers from memory while the page still says sourced |
| Re-validation of every live citation | A citation that slipped past the search provider's own domain filter |
| A mechanical check, with no AI call | Quoted passages absent from the retrieved material, and hosts imitating a government address |
| Cache discipline | A stale answer served after the law underneath it moved |
| Honest outage reporting | A provider failure papered over with an answer that reads as normal |
| Provenance stored with the answer | Old research reopening under a friendlier caption than it earned |
Because the caption is written in code rather than chosen by a model, the platform never shows a bare badge claiming an answer has been verified. It names only what was examined, and an answer with nothing quotable in it says so rather than borrowing reassuring wording it did not earn.
The limits
What the check does not cover.
A statement about what has been checked is worth nothing without a statement of what has not. Here is the second half, in the same words the app uses beside your answer.
The quoted passages and the links to government websites in a sourced answer, matched against the material actually retrieved for that answer.
Whether the analysis is legally correct, complete or current. Any citation to a statute, a regulation or a case. Any link that is not to a government website. Anything that is not a direct quotation.
A checked answer is a sourced answer whose quotations and government links were matched. It is not a verified statement of the law. That list is not copy written for this page: it is held as a named constant in the code, and it is the same list the product shows you.
See it for yourself
Ask it something you already know the answer to.
That is the honest way to judge any of this. Put a question to it where you already know how it comes out, set both panels to the same engine, and watch what retrieval and checking actually change.
Dickie Woods is a research and education tool. It gives general information about U.S. immigration law, not legal advice, and using it does not make Lemmy Law your lawyer. Immigration law changes often and an answer can be incomplete or out of date. Check important points against the sources cited, and speak to a qualified attorney about your own situation. Operated by Lemmy Law Pte. Ltd., a U.S. immigration law practice registered with Singapore's Ministry of Law as a Foreign Law Practice. Lemuel Lim is admitted in Missouri.