task_name: citation_extraction
manifest_version: "1.0.0"
ars_min_version: "3.10.0"
task_type: outcome-gradable
outcome_gradable: true
description: |
  Phase 1 gold subset for #182 verification_gate.verify_citation.
  Measures the lookup_verified 3-class enum classification accuracy
  across 51 hand-curated tuples. One of them (the OQ-5 by-design
  false-negative: a no-identifier fabrication) exercises the C-V6(a)
  coverage-gap path: a title-only unmatched stays unresolvable, never
  false, because no resolvable identifier exists to key the lookup on.
  The complementary real-but-unindexed canary (a verified genuine paper
  unmatched across all four resolvers) remains unfilled — see issue #250.
target:
  entrypoint: verification_gate.verify_citation
  predicted_field: lookup_verified
  expected_outcomes_path: expected_outcomes.json
  tuple_dir: tuples
sample_n: 51
labels: ["true", "false", "unresolvable"]
thresholds:
  aggregate:
    metric: accuracy
    direction: higher_is_better
    comparison: ">="
    threshold_value: 0.90
  per_class:
    metric: accuracy
    direction: higher_is_better
    comparison: ">="
    threshold_value: 0.85
    classes: ["true", "false", "unresolvable"]
tuple_distribution:
  - kind: valid_doi
    n: 20
    expected_lookup_verified: "true"
  - kind: valid_arxiv
    n: 10
    expected_lookup_verified: "true"
  - kind: manual_exempt
    n: 5
    expected_lookup_verified: "unresolvable"
  - kind: fabricated
    n: 15
    expected_lookup_verified: "false"
  - kind: fabricated_title_only
    n: 1
    expected_lookup_verified: "unresolvable"