Is the DOI retraction lookup free?

Yes. GET https://pay.ashesborn.cloud/api/v1/judge/paper?doi=… returns paper-level retraction status with no key and no account. The propagation and evidence-audit judgments are the paid endpoints.

What does an abstention mean?

The paper is outside the current snapshot, so the service declines to judge it instead of reporting “not retracted”. The reason is returned in the abstention_reason field.

Can I use the dataset commercially?

The integrity graph is published under CC0-1.0, so no attribution is legally required, though a link back is appreciated. The synthetic smart-contract corpus is MIT.

How current is the data?

The published snapshot is ri-20261025-1526, built on 2026-10-25. Every response carries the snapshot that produced it, and retraction dates are reported per source.

Datasets

Data with judgments, not just documents

Retrieval returns passages. These datasets return decisions you can check: a paper is retracted or it is not, a claim rests on retracted work or it does not, and every answer names the source row it came from. Start with the free DOI lookup below — it runs against the live snapshot and needs no key, no account and no model.

How the integrity graph works Measured benchmark

Free · no key · no account

Check any DOI for retraction

One HTTP GET against the live snapshot. The response lists every source that covers the paper, its source tier (A = registry-grade, C = index-derived) and the retraction date when a source records one. Nothing is generated by a model, so the same DOI returns the same answer.

The three datasets

Graph dataset

Citation + retraction integrity graph

Three tables with stable IDs: papers, citations, and one row per (paper, source) retraction assertion with a source tier. Enough to rebuild the citation neighbourhood of a retracted paper offline, or to audit how a judgment was reached.

Papers (nodes): 98,868
Citation edges: 144,814
Retraction status rows: 572
Snapshot: ri-20261025-1526

Sources: OpenAlex works and references, Crossref update-to relations, Retraction Watch (reason and nature free text excluded).

CC0-1.0Free to download

Hugging Face →

Judgment API

Retraction propagation judgment

Give it a DOI, get back whether its claims rest on retracted work: verdict (CLEAR / CONTESTED / …), confidence with its basis, reference coverage, retracted roots with depth and per-source agreement, conflict flags, and evidence-path length. No model in the loop, so two calls with the same snapshot agree.

Output: verdict + confidence
Per call: deterministic, 0 LLM tokens
Evidence: retracted roots, depth, agreement

Sources: Same snapshot as the graph dataset.

CC0-1.0 data, hosted judgment$0.02 per call

GitHub →

Synthetic corpus

Synthetic smart-contract security relations

A worked example of what a curated “data with judgments” delivery looks like: source-grounded JSONL, a PostgreSQL load script, and a Markdown corpus you can index with the GraphRAG MCP server. Use it to evaluate the pipeline without touching your own documents.

Entities: 11
Relations: 13
Source excerpts: 5
Evaluation questions: 5

Sources: Original fictional audit corpus shipped with graphrag-ts. Every contract, dependency and incident is invented.

MIT$0.02 local / $0.10 global query

Source →

What you can and cannot conclude

An abstention is an answer. When a DOI is outside the snapshot, the response says so (abstention_reason: paper_not_in_snapshot) instead of guessing “clean”.
“No status rows” is not “not retracted”. A paper can be in the graph with no source asserting either way. The dataset page shows both cases side by side.
The public graph is a neighbourhood sample. It is 12 retracted seeds with up to 60 citing works each plus their reference lists — not the whole scholarly record.
Source tiering is explicit. Tier A means a registry that records retractions directly (Crossref update-to, Retraction Watch); tier C means an index-derived flag (OpenAlex). A judgment that mixes them shows the mixture.
No reason text. Retraction Watch reason and nature free text are deliberately excluded, so a judgment never turns a retraction into an accusation of misconduct by paraphrase.

Endpoints

GET https://pay.ashesborn.cloud/api/v1/judge/paper?doi={doi}

Free, no auth, CORS open. Paper-level status from every covering source.

GET https://pay.ashesborn.cloud/api/v1/datasets

Machine-readable catalogue: IDs, licenses, pricing, tags.

POST https://pay.ashesborn.cloud/api/v1/judge

Paid per call through the x402 flow ($0.02 propagation, $0.10 evidence audit); the challenge arrives in the payment-required header.

https://mcp.ashesborn.cloud/mcp

MCP server (Streamable HTTP). Anonymous read tier: integrity_paper_status answers with no credential, so a client can call it the moment it connects. retraction_propagation ($0.02) and integrity_evidence_audit ($0.10) are paid calls and return payment_required to an anonymous session; indexing and the corpus-query tools need a credential.

Connect an MCP client

Points at the same snapshot the pages above use. No key for the free status tool; the server answers initialize and tools/list anonymously so a client can discover it first and decide about the paid calls later.

{
  "mcpServers": {
    "retraction-integrity": {
      "type": "http",
      "url": "https://mcp.ashesborn.cloud/mcp"
    }
  }
}

Call integrity_paper_status with {"doi":"10.1016/S0140-6736(97)11096-0"} — the same shape the browser lookup returns. Free, no key.
retraction_propagation and integrity_evidence_audit answer payment_required without a credential. Pay per call through POST https://pay.ashesborn.cloud/api/v1/judge (x402 on Base) or bring a credential.
Sessions expire; re-initialize rather than reusing a stale mcp-session-id.