Technology · The citation firewall
§ The problem
A general-purpose language model asked about Guam law will answer confidently — and will, sooner or later, invent a case. It will give you a citation in the right shape, a holding that sounds right, and a year that fits. None of it need be real. Courts have already sanctioned lawyers for filing briefs built on these fabrications. For a body of law as thinly documented as the territories', where a reader often cannot easily check the reporter, a plausible-sounding hallucination is worse than no answer at all.
Territorial Review takes the opposite stance. We use AI to read and synthesize the law, but we never let it be the authority on whether a case exists or says what it is claimed to say. That job belongs to the corpus. The model proposes; the database disposes. We call the mechanism that enforces this the citation firewall.
§ Three gates
When you ask a question, we first run a real search over the corpus and pull the opinions that actually bear on it. Only those passages — never the model's training memory — are handed over as raw material, and the answer comes back broken into segments: one clause at a time, each tagged with the citations that support exactly that clause. Every one of those citations is then put through three independent checks before it ever reaches your screen.
Each territory's reporter has its own citation grammar. A genuine Supreme Court of Guam cite takes the form YYYY Guam N. We parse every citation against that pattern; anything that does not match — a mainland federal cite, a state case, a citation borrowed from another jurisdiction — is flagged out of jurisdiction. This catches the cross-pollination error, where a model answers a Guam question with California law because that is what it saw most during training.
A citation can be perfectly well-formed and still be fiction. So we check each cite against the actual opinions retrieved for your question. If the model offers a cite in the right shape that does not correspond to a real opinion in front of it, we treat it as a possible fabrication and mark it unverified. The model is not allowed to vouch for the existence of a case. Only the database can.
The hardest failure to catch is the real case cited for a proposition it never stood for. For every citation that clears the first two checks, a separate model pass compares the statement against the actual text of the cited passage and asks a strict question: does this passage genuinely support this clause? If not, the citation is downgraded to unsupported, even though the case is real and correctly formatted. A verified mark means all three checks passed.
§ What you see
The result is rendered honestly. Each synthesized clause carries its citations as small badges, color-coded by what the firewall found, with a summary line telling you how many survived in each category:
Citation firewall✓ verified⚠ unverified / unsupported✗ out-of-jurisdiction
A green check means the cited opinion exists in the corpus and its text actually supports the statement — and the badge is a live link straight to the supporting paragraph. A warning means the cite could not be confirmed, or the passage does not appear to back the claim. A red mark means the citation does not belong to this jurisdiction at all. Nothing is hidden; the model's reach is shown alongside its grasp.
§ The same discipline, everywhere
The firewall is not one feature on one screen. The same posture — the model may read and explain, but only the corpus certifies what is real — governs every place AI touches the law.
§ Audited, not asserted
A safeguard you cannot measure is a slogan. So the firewall is evaluated, not just trusted: an offline harness runs a fixed set of questions twice — once with the guardrail and once without — and counts how many ungrounded or out-of-jurisdiction citations the firewall removes. That cross-pollination number is the firewall's reason for existing, made into something you can watch go up.
The whole apparatus is also jurisdiction-as-configuration. The citation grammar, the issuing court, the reception rule, and the persuasive-authority hierarchy live in a single block of settings. Pointing the same firewall at the CNMI, the U.S. Virgin Islands, or Puerto Rico is a matter of swapping that configuration, not rebuilding the verification logic. Our method is AI-forward; our standard is human-accountable.