What it verifies

A citation is extracted with regular expressions — never a language model — and then looked up in three tiers: a cache of previously confirmed results, a local corpus of court opinions, and finally the CourtListener API. The verdict is the outcome of that lookup.

No language model sits anywhere in the verification path. That is the whole point: the same citation submitted twice returns the same verdict, and no model can talk the engine out of one. An answer here is reproducible rather than generated.

What it does not do: judge whether a real case supports the proposition it was cited for. It verifies existence and identity, not reasoning.

The three verdicts

Price, networks and the caching rule

Route	Price	Notes
POST /api/v1/agent/verify	$0.01	One citation.
POST /api/v1/agent/verify/bulk	$0.10	Up to 50 citations, one flat price for the call. Fifty citations cost the same as two.
MCP tool verify_citation_paid	$0.01	Same answer over MCP at https://veridictum.legal/mcp/agent.

Payable on either chain, in USDC:

• Base — eip155:8453
• Solana — solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp

Both are advertised in the 402; pick either. The receiving address for each is in the accepts list of the Payment-Required header — never hardcode one from this page.

Request and response

Request:

POST /api/v1/agent/verify
Content-Type: application/json
PAYMENT-SIGNATURE: <base64url of the signed x402 payload>

{"citation": "Miranda v. Arizona, 384 U.S. 436 (1966)"}

Response — identical to the free, keyed route's envelope:

200 OK
PAYMENT-RESPONSE: <base64url settlement receipt: success, tx, network, payer>
X-Veridictum-Surface: agent

{
  "citation": "Miranda v. Arizona, 384 U.S. 436 (1966)",
  "status": "verified",
  "verified": true,
  "case_name": "Miranda v. Arizona",
  "court": "Supreme Court of the United States",
  "date_filed": "1966-06-13",
  "courtlistener_url": "https://www.courtlistener.com/opinion/107252/miranda-v-arizona/",
  "courtlistener_id": "107252",
  "confidence": "high",
  "tier": 1,
  "message": "Citation verified",
  "source": "local_corpus",
  "cached": true,
  "api_call_made": false,
  "name_match": "confirmed"
}

The bulk route takes and returns:

POST /api/v1/agent/verify/bulk
{"citations": ["384 U.S. 436", "Varghese v. China Southern Airlines Co., 925 F.3d 1339 (11th Cir. 2019)"]}

{
  "success": true,
  "total_citations": 2,
  "verified_count": 1,
  "not_found_count": 1,
  "suspicious_count": 0,
  "results": [ /* one envelope per citation, in order */ ],
  "verification_time_ms": 412,
  "hallucination_risk": "MEDIUM"
}

hallucination_risk is LOW when every citation verified, MEDIUM at one or two not found, and HIGH at three or more.

The 402 flow, in five lines

1. POST the route with no payment header. You get 402.
2. The requirements are in the Payment-Required response header, base64url-encoded JSON — not in the body, which is {}.
3. Decode it, pick one entry from accepts (one per network), and sign a payment for that exact amount to that exact payTo.
4. POST the same request again with the signed payload in the PAYMENT-SIGNATURE header.
5. You get 200 and the verdict, plus a settlement receipt in the PAYMENT-RESPONSE header.

Any x402 client library does steps 2 to 4 for you. Nothing is charged for the unpaid 402, and nothing is settled if the handler fails — the payment settles after a successful answer, not before.

Over MCP

The same verification is a paid MCP tool, verify_citation_paid, at https://veridictum.legal/mcp/agent (streamable HTTP), at $0.01 per call. The payment rides in the tool call's _meta under x402/payment; the settlement receipt comes back in the result's _meta under x402/payment-response. An unpaid call returns the payment requirements instead of a verdict.

The paid endpoint exists for agents that have a wallet and no account. If you have a Veridictum account and an interactive client, the free tools at https://veridictum.legal/mcp are the ones you want.

Verify the numbers yourself

Accuracy claims about citation verification are easy to make and rarely checked. These two are checkable, and the recipe is below — you do not have to take the figure on trust, and you should not.

How to re-run it. Submit any citation to the paid route with its case name attached, and compare the courtlistener_url in the response against CourtListener directly. Every verified answer carries the opinion it resolved to, so any verdict here can be checked against a source we do not control. To reproduce the mismatch behaviour, submit a real reporter citation with the wrong case name and confirm the reply is suspicious / mismatch rather than verified.

How the figures were arrived at is written up in docs/METHOD.md (the audit, the eight parser defects it found, and why the test strings are taken from the corpus rather than written by hand) and docs/REPORTER_COVERAGE.md (reporter coverage, and the four occasions on which something adjacent to the question was measured and reported as the question). Both live in the project repository.

One caveat stated plainly, because it is the honest version of the claim: this verifies that a citation exists and that its name matches the case it points to. It does not verify that the case supports the argument it was cited for. Substance still needs a lawyer.

Also

• /llms.txt — the short machine-readable version of this page.
• A free, API-key-authenticated surface exists at /api/v1/public/* with the same envelope, for callers that would rather hold a key than a wallet.
• Terms of Service — section 11 covers API, MCP and pay-per-call use. There is no SLA: a settled call buys that call, not capacity, priority or future availability.
• Privacy Policy — an x402 call carries no account. We record the settlement (timestamp, route, network, payer wallet, amount, transaction hash) and nothing about the caller.