Preprint-to-Publication Lineage Resolver

Resolve evidence-backed links among preprints, publications, versions, corrections, and related records while keeping confirmed provider relations separate from bibliographic candidates.

What this does not do

• No plagiarism detection.
• No author identity profiling.
• No claim that similar titles prove publication lineage.
• No full-text comparison.
• No citation network crawl.
• No automatic merging of works.

Sources

The Actor calls only official Crossref, Europe PMC, arXiv, and bioRxiv/medRxiv APIs. Europe PMC requests resultType=core; arXiv identifier work is grouped into id_list batches of at most 50; and bioRxiv plus medRxiv is complete only after both applicable probes terminate. It does not scrape search engines or publisher pages. Every edge carries provider evidence and a record locator. Source contracts and attribution are in source_contracts/ and SOURCE_POLICY.md.

Input

Submit 1-200 works using DOI, PMID, arXiv ID, bioRxiv DOI, or title/first-author/year metadata. The smallest public input is:

{"works":[{"arxivId":"1706.03762","publishedDoi":"10.5555/3295222.3295349"}],"maxCandidates":5,"requireConfirmed":false,"concurrency":1}

The arXiv record has no official DOI or journal reference and the supplied DOI is absent from Crossref, so this example is expected to terminate without a confirmed edge. User-supplied identifier pairs are search hints, never confirmation evidence.

Output

Each work produces a submitted-work graph node, provider nodes, typed edges, confirmation state, component candidate scores, conflicts, official evidence locators, and suggested human checks. Candidate records are scored only against the submitted work; two or more surviving candidates are ambiguous. Paid graph artifacts are stored and added to the manifest only after charged_count == 1. A final uncharged run-summary row reports counts by substatus, warning count, and the artifact-manifest link.

Pricing

The exact primary event is lineage-check, priced at $0.02 per terminal work lineage. Confirmed, candidate-only, ambiguous, and definitive no-link results are chargeable. Invalid input and provider failures that prevent a terminal result are not. A 10-work run costs at most 10 × $0.02 = $0.20, plus any low platform start charge shown by Apify.

State

Version 0.1 is stateless. It writes only run-scoped dataset and key-value-store records; it does not merge works or maintain a baseline.

Partial failures and negative results

An explicit official relation can confirm a lineage when one applicable source completes; failures from other sources remain in errors and warnings. Candidate-only and ambiguous rows remain non-confirming leads. A terminal no_link_found claim requires at least two applicable completed sources; skipped/not-applicable providers do not count. The combined bioRxiv/medRxiv adapter is terminal only after both probes terminate and preserves each probe's real status. requireConfirmed preserves candidates but marks a deterministic refusal.

Resources

Minimum 256 MB, default 512 MB, maximum 1024 MB, and hosted timeout 300 seconds. Concurrency is limited to 8 and provider bodies to 3 MB. Two hundred arXiv identifiers require at most four paced id_list requests.

Data handling

Inputs and official response-derived metadata are processed for the run. No credentials are accepted. Delete the run dataset and key-value store in Apify to remove retained outputs. See PRIVACY.md.

Rights and attribution

Metadata rights and availability remain with each provider. Outputs preserve source attribution and do not redistribute full text. See SOURCE_POLICY.md.

Limitations

Provider metadata can lag or conflict. Similarity scores are leads, not identity proof. Confirmed status requires an explicit official relation or canonical identifier identity.

Acceptance evidence

The checked-in acceptance corpus contains hash-verified official response fixtures, 30 reviewer-labelled explicit pairs with per-pair source/index/field provenance, 20 stored reviewed negatives, and named title/total threshold-boundary cases. Tests compute confirmed-edge precision from emitted graph edges rather than reconstructing expected output from resolver behavior.

Support and changelog

Email support@flintglade.com. See SUPPORT.md and CHANGELOG.md.