schema_version: 1
benchmark_release: argo-audit-2026-07-11
baseline_commit: f661dc2298c040d22ef8cf070c349d32616e2073

cohorts:
  argo139_rl_narrative:
    role: collected_cohort
    expected_members: 139
    membership_file: genes.csv
    prediction_suffix: -bioreason-rl-predictions.md
    review_suffix: -bioreason-rl-review.md
    model_source: app.bioreason.net/RL
    model_version: unavailable_in_web_export
    export_timestamp_source: per-file BioReason export header
    frozen_inputs:
      goa: per-gene committed TSV at baseline_commit, identified by SHA-256
      ontology_release: "2026-03-25"
      ontology_url: https://release.geneontology.org/2026-03-25/ontology/go-basic.obo
      ontology_sha256: a77e356737dab39a4f620dce35fc4d6eb531c4b6153af6cacaaa322b49b804bd
      ontology_integrity_sentinels:
        obsolete: [GO:0000002, GO:0000003, GO:0005615, GO:0005844, GO:0007568]
        active: [GO:0005576, GO:0022414, GO:0051082]
      ontology_external_verification:
        checked_on: "2026-07-12"
        command: uv run python projects/BIOREASON_COMPARISON/verify_ontology_authority.py
        official_archive_sha256_match: true
        live_obsolete_terms: [GO:0000002, GO:0000003, GO:0005615, GO:0005844, GO:0007568]
        authorities:
          - https://www.ebi.ac.uk/QuickGO/services/ontology/go/terms/{GO_ID}
          - https://www.ebi.ac.uk/ols4/api/ontologies/go/terms?obo_id={GO_ID}
      model_input_sections: >-
        The full sequence, InterPro Domains section, and upstream GO-GPT GO Terms
        section are preserved in each committed raw web export. Per-section counts
        and SHA-256 checksums are recorded in benchmark-quality.csv.
    reference:
      source: local AIGR gene review
      interpretation: agent-adjudicated local reference, not independent expert ground truth
      status_field: status
    scoring:
      minimum: 1
      maximum: 5
      axes:
        correctness: factual support for claims that the model actually makes
        completeness: coverage of established core functions, locations, complexes, and pathway context
    input_quality:
      model_sequence_limit: 2000
      truncated_inputs_are_stratified: true
      truncated_inputs_are_excluded: false
    performance_exclusions:
      - species: worm
        symbol: csr-1
        reason: WRONG_INPUT_SEQUENCE
        details: >-
          The benchmark requested CSR-1 (Q21992), but the cached UniProt record and
          BioReason input are nhr-47 (Q17370). The resulting nuclear-receptor summary
          evaluates the supplied sequence correctly and is not a model-function error.

  argo95_sft_terms:
    role: primary_sft_term_cohort
    expected_members: 95
    parent: argo139_rl_narrative
    model_source: wanglab/protein_catalogue
    novelty_rule: >-
      COR means absent from the committed per-gene GOA snapshot; an exact GOA
      identifier match is CNN unless a separately documented temporal rule applies.
    ontology_status_rule: >-
      Record ontology status in every affected review rationale and include every
      mismatched or unresolved pair that is biologically nonnegative at the audit
      baseline or after reclassification in the ontology-pair audit. Do not change the
      biological assessment solely because of ontology status. CNN, COR, UNC, and LSP
      retain their schema meanings; LSP is used only when the supplied ID denotes a more
      generic concept than the supported annotation, not as a catch-all for a stale
      identifier whose historical concept is unchanged.

reporting:
  distinguish_collection_and_performance_denominators: true
  report_reference_status_distribution: true
  report_input_quality_distribution: true
  generated_sidecars:
    - benchmark-cohorts.csv
    - benchmark-genes.csv
    - benchmark-quality.csv
    - benchmark-metrics.json
    - second-review-agreement.json