---
name: best-of-evidence
description: Rank competing AI answers by citation coverage, source authority, claim support, and verifiability. Use when choosing among sourced answers, research summaries, or evidence-backed recommendations.
---

# Best of Evidence

## Definition

Best of Evidence is an evidence-guided selection approach that ranks candidate answers by the quality, coverage, and verifiability of their supporting sources. This skill adapts the research term to citation assessment, rewarding checkable claims and primary sources rather than confidence or writing style alone.

Reviewed 26 July 2026.

Choose the answer supported by the strongest checkable evidence, not the most confident prose.

## Workflow

1. Split each candidate answer into important factual claims.
2. Map citations to the claims they directly support.
3. Score coverage: supported important claims divided by all important claims.
4. Score authority: prefer primary sources, official documentation, standards, and original research.
5. Check whether citations actually entail the associated claims.
6. Penalize missing dates, inaccessible sources, citation laundering, and sources that merely mention the topic.
7. Rank candidates and explain the decisive evidence gap.

## Output

Return a ranked list, score out of 100, unsupported claims, strongest sources, and a concise selection rationale.

Do not treat citation count as evidence quality.

## Research basis

[Best-of-Evidence: Best-of-N Selection under Partial Verification](https://arxiv.org/abs/2607.20950) — Zhang et al., arXiv (2026). The paper introduces the exact term for evidence-guided candidate selection; this skill adapts it to citation assessment.