For AI legal-tech vendors

The citation-verification API for AI legal tools.

Your users ask "will it hallucinate cases?" — end the objection. Every citation your product outputs, checked against the actual reporter. Deterministic. $0.05/citation, usage-billed.

Get an API key Read the docs

Who uses this

AI legal-drafting startups (the "Cursor for lawyers" wave), AI legal-research and brief-summarization products, contract-analysis tools that surface case law, and any product where a user might paste an LLM-generated citation into a court filing. If your customers are lawyers, and your product ever outputs a case citation, that citation needs to be verified before it ships. That's our whole job.

The math

A typical AI-drafting product priced at $199/mo per lawyer generates ~30 briefs per month per seat, averaging 15 citations per brief. That's 450 verification checks per seat per month — $22.50 at our first-tier price. That's ~11% of your MRR passed through as verification cost, in exchange for the single largest objection your prospects raise before buying evaporating.

Above 10,000 checks per month per key, the price drops to $0.03/citation.

Reference implementation

Single citation:

curl -X POST https://citationsafe.com/api/v1/verify/citation \
  -H "Authorization: Bearer cs_live_..." \
  -H "Content-Type: application/json" \
  -d '{"citation": "Brown v. Board of Education, 347 U.S. 483 (1954)"}'

Whole document (extracts every citation):

curl -X POST https://citationsafe.com/api/v1/verify/document \
  -H "Authorization: Bearer cs_live_..." \
  -H "Content-Type: application/json" \
  -d '{"text": "The court held that separate is inherently unequal. Brown v. Board of Education, 347 U.S. 483 (1954). See also Miranda v. Arizona, 384 U.S. 436 (1966)."}'

Full response shapes, error codes, and rate-limit headers in the API reference.

Pricing

$0.05 per citation for the first 10,000 checks per calendar month, per API key.
$0.03 per citation above 10,000 checks per calendar month.
Set a hard monthly cap to prevent runaway bills. Overage billing is opt-in per key.
No monthly minimum, no per-seat charge, no setup fee.

Get an API key

MVP: we manually issue keys after a brief eyeball. Fill this in and we'll be back inside a business day.

SLA + terms

Target uptime 99.5% (best-effort during MVP), citation-verdict correctness backed by our $10,000 fabrication warranty — the same guarantee we offer web-app subscribers. See the vendor-API terms of service for the full contract.

How teams integrate

Pattern 1: pre-output guard

The generative model produces a draft; every citation in the draft is checked before the draft is returned to the user. Failed citations are either stripped, flagged inline, or the whole draft is regenerated with a “do not cite” instruction added. This is the highest-safety integration and the one that maps most cleanly to your “does not hallucinate” marketing claim.

Pattern 2: badge-on-output

The citation ships to the user regardless of check result, but each citation carries a visible VERIFIED / UNCONFIRMED / NOT FOUND badge next to it. Cheaper (checks can run async) and lower-risk to your existing UX, but leaves the user responsible for reading the badge.

Pattern 3: export-time gate

Free-form generation continues unchanged; the check runs only when the user hits “export to Word / PDF / clipboard.” Cheapest and lowest-friction — you only pay for the checks that gate a real filing risk — but only works if your product has a discrete export step.

What comes back

Per-citation response includes a verdict (VERIFIED / UNCONFIRMED / NOT FOUND), the primary source the citation was matched against, a case-name normalization, and a per-layer breakdown (existence, quote-match if quoted text supplied). Latency is typically under 800ms for a cached lookup and under 4s for a first-time lookup that requires an external fetch. Rate-limit headers on every response tell you your current per-key quota and reset window. See the API reference for the full JSON schema, error codes, and idempotency semantics.

Frequently asked

Is the check deterministic or LLM-based?: Deterministic. Every layer is a database lookup or a fuzzy-string match against actual opinion text — no LLM in the loop for verdicts. This is the entire product differentiator: an LLM-based verifier can hallucinate a “yes, this exists” the same way an LLM drafter hallucinates the citation in the first place. Ours can’t.
What’s the false-verify rate?: 0.000% on our published eval set. Backstopped by the $10,000 fabrication warranty — if an API response returned VERIFIED for a citation that turns out to be fabricated, we pay. Full methodology at /quality.
How do you handle PII in the payloads?: Payloads are processed in-memory and not persisted beyond an anonymous request-count row for rate-limiting. See our privacy policy and the DPA we’ll sign under vendor contract.
Do you support co-marketing?: Yes — a “Verified by Citation Safe” badge asset and case-study collaboration are available on request. Ask when you fill out the API-key form below.
What jurisdictions do you cover?: All U.S. federal courts + all 50 states + DC published opinions; UK, Canada, Australia, NZ for entity-level existence checks. Statutes, regulations, and U.S. Tax Court opinions supported on the same endpoints. Full coverage detail at /coverage.
Can we self-host or is it SaaS-only?: SaaS-only today; self-hosted VPC deployment available on the Suite plan for six-figure ACVs. Note it in the API-key request form.
Rate limits?: Default 60 requests/second per API key with burst headroom. Bumps on request — no additional cost, we just need to know you’re coming so we can size capacity ahead of launch.
Do you have client libraries?: Official TypeScript and Python clients ship alongside the raw HTTP API. See /docs/api/v1 for install commands and usage examples.