Scholarly evidence graph
A persistent, queryable layer connecting works to citations, versions, repeat attempts, recorded outcomes, and screening runs without turning proximity into a verdict.
What it stores
Canonical work nodes are resolved through DOI, arXiv, and internal identifiers. Directed edges record cites, is_version_of, replicates, and reproduces relationships. Time-stamped observations record source facts such as FLoRA outcomes, problematic-paper dataset matches, recorded publication outcomes, and versioned preprints.ai screening runs.
Every edge and observation names its source and source record. The projector is idempotent, records phase counts, and refreshes the graph from the platform’s current data.
Public API
Graph coverage and freshness reports aggregate node, edge, observation, and source counts.
GET /v1/evidence-graph/{identifier} returns a bounded one-hop neighbourhood for a DOI, arXiv identifier, or platform paper ID. Report cards link to the same JSON.
Coverage
- All works and identifiers in the preprints.ai catalogue.
- Published-version links held in platform metadata.
- Replication and reproduction edges from FLoRA.
- Citation edges extracted from cached manuscript full text.
- Problematic-paper, recorded outcome, and current screening observations.
Known limits
DOI extraction is coverage-oriented and does not infer rhetorical citation context. Works without stable identifiers can remain separate nodes. Source coverage is uneven across disciplines, dates, and publication types. The public endpoint deliberately avoids an interactive force-directed visualisation because visual proximity can imply a strength or certainty the data does not contain.