Stage audit: bibliography hygiene checks per reference and writes audit_quality/audit_confidence/audit_remediation per row.
Stage ground: evaluates whether each reference supports the claims it is cited for; fetches evidence preferring md/html/tex/rtf/txt/pdf; writes ground_* columns and produces grounding reports.
Initialize the reference QA pipeline by creating out/audit_references.csv with one row per BibTeX entry and required stage columns.
Stage ml: ML venue lens (default NeurIPS) scoring per reference for technical fit: canonicality, baseline relevance, recency for SOTA claims; writes ml_* columns.
Stage resolve: resolve each reference against OpenAlex/Crossref/arXiv/DBLP; detect hallucinations and metadata mismatches; writes resolve_* columns and refs.corrected.bib.
Final stage: compute reference_quality_score by weighted QUALITY×CONFIDENCE across audit/resolve/ground/venue/ml; rank and assign review_priority; includes default + venue-specific blocker rules.
Stage venue: policy venue lens (governance/ethics/policy) scoring per reference based on authority/genre appropriateness for policy claims; writes venue_* columns.