How we establish trust

The graph makes a specific promise: it may contain weak claims, but it never conceals their weakness. Every citation edge carries an honest statement of how far to trust it, and where that trust comes from.

Three kinds of edge

Not every citation is the same kind of thing, so we never present them as if they were.

• Citation of record

  Citation of record

  Hand-verified cross-references. Authoritative. This is what the citation surface returns.

• Machine-supported

  Machine-supported reference

  An explicit citation whose evidence is verified against the canonical text at both ends. Useful and inspectable — and explicitly machine-derived.

• Research lead

  Research lead

  A proposed relationship that has not been verified to citation-of-record standard. A direction to investigate, never presented as settled.

The API keeps these separate by construction. The strict citation surface returns only citations of record. Machine-supported references and research leads are available too — always labeled as what they are.

Confidence is earned, never asserted

We do not stamp certainty on machine inference. A confidence figure is meaningful only when it rests on one of two things: recognized authority (a hand-verified source, or a human who reviewed the exact claim) or calibration against human-ratified ground truth.

An extraction that merely looks confident earns nothing. Neither does surface similarity, nor two systems agreeing. Those are signals; only calibration or authority turns a signal into a confidence.

Where no legitimate basis exists yet, an edge reads as unscored — which is more honest than a fabricated number.

Verified at both ends

For explicit citations, "machine-supported" means something precise and checkable: the cited words are found in the source at their stated location, and they map specifically to the claimed target under a stated rule — distinctive enough that they could not point somewhere else by coincidence. A generic phrase that happens to appear in many places does not qualify, no matter how confident the extraction sounded.

The human is sovereign

A recognized reviewer can affirm, correct, or reject any edge. A confirmation makes that exact edge authoritative. A correction creates a new, corrected edge and preserves the original as an auditable record — nothing is silently overwritten, and confidence is allowed to go down when the evidence says so.

Machine estimates get better over time as verification re-runs and as human rulings accumulate. But only a human ruling makes a claim authoritative — the machine proposes; the scholar decides.

What this means for you

Every edge you receive tells you its trust class, the basis for its confidence, and — where it applies — the exact words that ground it. You never have to guess whether a link is a settled citation or a promising lead. The graph tells you, on every edge, every time.