# Transparent methodology

## Scope

EvidenceGraph organizes **research records**, **documented passages**, **claims**, **descriptive indicators**, and **policy hypotheses**. It is not a machine learning classifier and does not generate conclusions from bibliographic metadata.

## Evidence linkage

Each directed link stores a source node ID, a target node ID, a relation label, an excerpt, a precise locator, a review status, and creation date. Relation categories are `supports`, `challenges`, and `context` and reflect a researcher's interpretation, not scientific verification.

## Documentation completeness

`located-excerpt coverage = (number of links with non-empty excerpt AND locator) / number of all links × 100`

When there are no links, coverage displays zero rather than a misleading 100%. A link can be completely documented yet factually wrong. Therefore this metric does not measure evidence strength, accuracy, independence or rigor.

## Mixed viewpoints / warnings

A claim with at least one `supports` link and one `challenges` link is flagged for review, **not** deemed logically inconsistent or scientifically disproven. A claim/hypothesis without any located excerpt is flagged as under-documented. Nodes without incoming or outgoing links are marked unconnected.

## Source variety

The count of distinct upstream nodes is only a registry count. It is **not** a count of independent studies: publication duplication can occur across titles or versions, and multiple edges can describe the same study.

## Path search

Breadth-first search finds the shortest connected sequence between two nodes, treating links as undirected for exploration. The resulting path has no causal or entailment meaning and is not a recommendation.

## Quantitative indicators

The World Bank import selects the latest non-null, numeric annual value from returned observations, retaining the original year and unit. Results must not be interpreted as contemporaneous if years differ. Indicators alone cannot validate a causal mechanism or establish treatment effects.

## Limits of evidence extraction

External APIs return **metadata** or numeric observations, not full-text evidence. Researchers must read primary sources and copy an authentic relevant passage with a page, section, table or stable locator. The app does not check excerpt authenticity. Use research protocols and independent review for publication-quality findings.