Data with judgments you can check

Three tables, one snapshot, stable IDs. papers (98,868 rows: OpenAlex work ID, DOI, title, year) · cites (144,814 directed edges, each with the source that supplied it) · retraction_status (572 rows, one per paper/source assertion, with a source tier and the date that source records).

Keeping assertions separate is the design decision that matters. A retraction is not one boolean from one authority; it is a set of claims that can disagree, and collapsing them into one is_retracted column destroys the only thing a reader can audit. Retraction Watch reason and nature free text is excluded on purpose — carrying a publisher's prose turns a bibliographic lookup into an accusation.

The lookup answers “is this paper retracted, according to whom”. The paid calls answer the harder question — “does this paper's argument rest on retracted work” — and are priced per call through x402 on Base, so a program can pay without a subscription or a sales conversation.

GET https://pay.ashesborn.cloud/api/v1/judge/paper?doi=10.1016/S0140-6736(97)11096-0

{
  "paper": { "paper_id": "W2117847125", "doi": "10.1016/s0140-6736(97)11096-0", "year": 1998, "source_tier": "C" },
  "statuses": [
    { "source": "openalex",         "source_tier": "C", "is_retracted": true, "retraction_date": null },
    { "source": "retraction_watch", "source_tier": "A", "is_retracted": true, "retraction_date": "2010-02-06" }
  ]
}

THREE ANSWERS PEOPLE MISREADAn abstention (paper_not_in_snapshot) means the service declines to judge — not that the paper is clean. No status rows means the paper is in the graph but no source asserts anything — absence of evidence, a different verdict. Source disagreement is reported, not silently resolved in favour of the higher tier.

Twenty DOIs — 16 retracted records drawn from the status table, three controls, one invalid DOI — queried 2026-10-28 with one GET each, no retries: 20/20 HTTP 200, median 1,595 ms, slowest 5,318 ms (cold start), and all 16 retractions confirmed by three sources. The engine itself reports single-digit milliseconds in server-timing; the rest is network.

Full benchmark, method and the reproducible loop ↗ · Schema walkthrough and the rendered citation graph ↗

The public subset is a neighbourhood sample: 12 retracted seeds, the works that cite them, and those works' reference lists — 781 citing papers, 788 edges, 6 papers citing two or more retracted seeds. It is not the scholarly record. Retraction dates are per source; a null means that source did not date it.

The MCP endpoint https://mcp.ashesborn.cloud/mcp answers anonymous initialize and tools/list and exposes one free read-only tool, integrity_paper_status, so a client can call it with no key. retraction_propagation ($0.02) and integrity_evidence_audit ($0.10) return payment_required to an anonymous session — the free tool is deliberately not a copy of the paid ones. Indexing and the corpus-query tools still need a credential. Sessions expire: re-initialize rather than reusing a stale mcp-session-id.