Evaluate the prototype on an expert-annotated gold set

[sudoantonio]

Objective

Measure the empirical quality of filtering, semantic extraction and association
against judgements produced by domain experts.

Acceptance criteria

• Corpus sampling, provenance and annotation instructions are documented.
• Metrics include class accuracy, field accuracy, evidence accuracy and abstention.
• Inter-annotator agreement and adjudication are reported.
• Results are reproducible and clearly separated from synthetic tests.