Trust

Where we pull sources

CiteSafe is the trust layer for legal AI. We do not ask the model that wrote the brief to grade the brief. Every status is backed by named layers — risk, corpus, and public primary indexes — with honest limits.

Last updated: July 2026 · Engine target: multi-source API (v3) · Scout offline fallback always available.

Verification pipeline

Extract → score risk → corpus → primary lookup → export. We do not draft briefs.

1
Extract
Pull every case, statute, rule, and incomplete cite from the draft.
2
Risk
Run fabrication heuristics before anything gets a green light.
3
Corpus
Match curated high-signal authorities (offline + API).
4
Primary indexes
CourtListener + CAP for cases; GovInfo signals for statutes.
5
Export
Verified · Needs Review · Could Not Verify + source lineage + report hash.

Source stack

What each layer proves — and what it deliberately does not.

Fabrication heuristics

Risk layer · Live in Scout

Proves: High-risk fake patterns: impossible years, invented reporters, fabricated party names.

Does not prove: Does not prove a real case exists or is good law.

CiteSafe authority corpus

Curated · Live in Scout

Proves: Match against a growing set of high-signal known authorities and common statute forms.

Does not prove: Not a full commercial library. Corpus match is not KeyCite or Shepard's.

CourtListener (Free Law Project)

Primary public index · API / Clerk+

Proves: Opinion/case search hit in a major open case-law index, with link when available.

Does not prove: Not negative treatment, quote accuracy, or holding support.

Open index

Caselaw Access Project (Harvard)

Primary public index · API / Clerk+

Proves: Case existence / citation match in CAP when the index returns a hit.

Does not prove: Not full commercial coverage; not treatment history.

Open index

GovInfo / public statute indexes

Public catalog · API / Clerk+

Proves: Signals and catalog links for U.S.C.-style statute forms.

Does not prove: Not full statutory research or currency of every section.

Open index

Citation form recognition

Structure · Live in Scout

Proves: Bluebook-ish structure detected (reporter, U.S.C., FRCP, Constitution).

Does not prove: Form alone is never existence.

Statuses (human in the loop)

Verified

Existence or high-confidence corpus/form match with named source.

Ambiguous (yellow)

Cite may be wrong or incomplete — open the linked primary source and prove it.

Could Not Verify

Fabrication risk or no index match — do not file.

We never market “hallucination-free.” Existence ≠ good law, holding support, or quote accuracy.

Scout · Clerk · Bench

Model	Depth	Status
Scout 1.0	Extract · corpus · heuristics · export	Live beta
Clerk 2.0	+ CourtListener · CAP · GovInfo · report hash	Live via Vercel API
Bench 3.0	+ bulk · automations · firm policy · API	Roadmap

Data handling

We do not use customer briefs to train a public generative model.
We do not sell your documents or personal data.
Scout beta history often stays device-local unless a server feature is enabled.
Primary lookups go through our API (not from the browser) when hosted.

Try a verification Watch demo Security

Third-party names (CourtListener, Free Law Project, Caselaw Access Project, Harvard, GovInfo, Westlaw, Lexis, and others) are for identification only. CiteSafe is not affiliated with or endorsed by those organizations. See Terms.

CiteSafe