Dicta is live. It provides hybrid search, grounded answers, and citation verification over the full public corpus of U.S. case law, built as an API that developers and AI agents can call directly.
The Corpus
- >10,326,074 court opinions, the full CourtListener public-domain corpus, in a full-text index
- >Federal statutes and regulations: the U.S. Code and the CFR
- >State statutes: Pennsylvania, Florida, California, New York, Delaware, and Ohio
- >18,060,118 resolvable citations in the verification index
Ranked by Authority, Not Keywords
Dicta blends BM25 full-text search with PageRank over the citation graph (8M cases and 77M citation edges), so landmark cases rise above incidental keyword matches.
As a sanity check, ranking purely on graph structure put *Miranda*, *Strickland v. Washington*, *M'Culloch v. Maryland*, *Gibbons v. Ogden*, and *Terry v. Ohio* near the top. The citation graph independently rediscovered the canon of American law.
Semantic vectors over opinion leads and chunked full text are fused with BM25, so a plain-English question finds what keyword search structurally can't.
Citation Verification
Lawyers have been sanctioned for filing briefs with citations an AI invented. Dicta's lead product is the guard against that:
- >Does it exist? Each citation is parsed and resolved against the corpus, then marked verified, unverified, or unparseable.
- >Is it still good law? Negative-treatment signals flag overruled or abrogated authority. Treatment extraction is direction-aware, so "X was overruled *by* Brown" no longer marks *Brown* as overruled. That single fix removed about 97% of false flags on landmark cases.
- >No language model on the hot path. Verification is deterministic.
API and MCP
- >Metered API at api.dicta.wire.band: API-key authentication, metered per verified citation.
- >MCP server with eight agent tools:
verify_citations,resolve_citation,search_caselaw,get_case,check_treatment,search_statutes,lookup_statute, andask_caselaw. - >Grounded answers cite every proposition as *(Case, year, opinion ID)*, using only retrieved sources. Repeat queries are answered from a semantic cache from the Wire.Band stack.
What It Cost to Build
- >Full-text index and citation graph: built in an afternoon on a CPU-only server we already owned, for $0 in infrastructure. The 10.3M-opinion index took 93 minutes, and PageRank took under a minute.
- >Full-text embeddings: about 4 hours on two rented RTX 5090s, for roughly $4.
Public-domain data, open-source search, and a few dollars of GPU time add up to a citation engine an agent can call. The expensive part of legal research turns out to be the ranking and the verification, and that's the part we built.