Evidence Grading

Usage Examples

Grade a specific paper

evidence_grading.grade_paper(
    doi="10.1038/s41586-024-00001-0",
    domain="machine_learning",  # affects grade rubric
    include_rationale=True
)
# Returns: {"grade": "B+", "level": "replicated_benchmark", "rationale": "..."}

Grade evidence for a claim across all supporting papers

evidence_grading.grade_claim(
    claim="Scaling laws hold for code generation models",
    supporting_papers=claim_tracker.get_papers("claim_0099"),
    schema="custom"  # strong | moderate | preliminary | anecdotal
)

Generate calibrated hedging language

evidence_grading.suggest_language(
    claim="Our method outperforms baselines on protein fitness prediction",
    grade="preliminary",
    context="manuscript_results_section"
)
# Returns: "Our method demonstrates promising performance improvements over baselines..."

Output Format

Returns grade, level label, confidence score, and rationale. For claim-level grading, returns aggregate grade with per-paper breakdown. Hedging language output is ready-to-use prose.

Best Paired With

• claim_auditor for calibrating claim strength
• experiment_skeptic when the paper's conclusion seems stronger than its evidence

Notes

• Grade rubrics are domain-aware: ML/CS standards differ from clinical/biomedical standards
• Preprints are graded conservatively until peer review is confirmed
• Combine with contradiction-detection — contradicted claims automatically receive a grade penalty