T·R

Technology · Territorial Review

From the scrape to the citation.

§ The system

Territorial Review is not a chatbot pointed at a search box. It is a built pipeline with a clear division of labor at every layer: the official record is collected and normalized, indexed for both keyword and meaning, read and synthesized by language models, and then — before anything reaches your screen — checked back against the corpus that produced it. The model proposes; the database disposes. Here is the whole stack, from the data it stands on to the page you read. Each layer is documented in full on the pages that follow.

06InterfaceIn-app opinion & statute reader · live in-text citations · pin-cite deep links · cite copier · ask-this-¶ · ⌘K palette
05VerificationCitation firewall (jurisdiction · existence · entailment) · deterministic statute-currency flags
04SynthesisPer-opinion holdings · the synthesized doctrine answer · statutes-first authority ordering
03RetrievalHybrid keyword + semantic search · Reciprocal Rank Fusion · legal-domain embeddings
02CorpusOpinions · the Guam Code Annotated · the legislative record — one normalized schema
01IngestionScrape the official sources · parse the PDFs · normalize · chunk · embed · coverage-verify

§ At a glance

One jurisdiction, loaded end to end. Guam is live today — every Supreme Court opinion, the entire code, and the legislative record that connects them.

792Supreme Court opinions1996–present
17,898Code sectionsAll 22 titles · ~3M words
3,222Public Laws22nd–38th Legislatures

§ Explore the build

§ Built to travel

None of this is wired to Guam by hand. The citation grammar, the issuing court, the reception rule, and the persuasive-authority hierarchy live in a single block of configuration. Pointing the same engine at the CNMI, the U.S. Virgin Islands, or Puerto Rico is a matter of swapping that configuration, not rebuilding the verification logic. As the corpus grows jurisdiction by jurisdiction, the firewall and the graph grow with it.

The discipline behind all of it is simple to state. Our method is AI-forward; our standard is human-accountable. The model accelerates the reading and the drafting. The corpus, the firewall, and a named editor decide what is allowed to stand.

§ Colophon

Built on a deliberately boring, durable stack — the interesting parts are the pipeline and the guardrails, not the plumbing.

Application
Next.js · TypeScript · deployed on Vercel
Data
PostgreSQL + pgvector on Supabase · row-level security · streaming responses
Embeddings
A legal-domain vector model · keyword and meaning searched in parallel
Language models
Frontier LLMs — for synthesis, one-sentence holdings, and citation-entailment checks; the model is never the authority on what is real
Ingestion
Python · PyMuPDF text extraction · coverage-verified loaders
See every layer at work on a live question.Go to the archive →