🎯 场景：审计某条 citation 是否真支持它声称的 claim（fully / partially / not supported）。RAG citation 质量评估的核心，能抓出"假引用"——citation 本身真实但内容不支持 claim。

Quick Use

Use when: You want to audit whether a citation actually supports the claim it was attached to. Fill in: {{claim}} = an atomic claim from the AI answer; {{cited_span}} = the verbatim source text the answer cited. You'll get: A fully/partially/not-supported verdict, the supporting phrase, and what's missing. Output is JSON.

Purpose

For a single (claim, cited_span) pair, decide whether the cited span actually supports the claim — fully, partially, or not at all. Used to audit citation faithfulness in RAG outputs at the per-claim level, so that "hallucinated citations" (a real citation pointing at unrelated text) can be caught even when the cited document is real. Produces a structured score per pair, which can be aggregated into citation precision metrics.

Prompt

You audit citations. Decide whether the cited span actually supports the
claim it was attached to.

Claim:
{{claim}}

Cited span (verbatim from the source document):
{{cited_span}}

Scoring rubric:
- "fully_supported"     — every part of the claim is directly entailed by
                          the span. No additional inference required.
- "partially_supported" — some parts of the claim are supported by the
                          span, but at least one part requires unsupported
                          inference or is missing from the span.
- "not_supported"       — the span does not entail the claim, or
                          contradicts it.

Be strict. If the claim adds a number, qualifier, or causal link that the
span does not state, that is at most "partially_supported".

Return ONLY this JSON object:
{
  "verdict": "fully_supported" | "partially_supported" | "not_supported",
  "supporting_phrase": "<verbatim sub-string of cited_span that supports the claim, or empty string if none>",
  "missing_or_extra": "<for partially/not_supported: what part of the claim is not in the span, in <=20 words; empty string for fully_supported>",
  "decision_basis": "<one sentence, <=30 words, no internal CoT>"
}

Example

Input:

claim: "The James Webb Space Telescope launched on December 25, 2021 and reached L2 in January 2022."
cited_span: "The James Webb Space Telescope launched on December 25, 2021, on an Ariane 5 rocket from French Guiana."

Expected output:

{
  "verdict": "partially_supported",
  "supporting_phrase": "The James Webb Space Telescope launched on December 25, 2021",
  "missing_or_extra": "The span does not mention reaching L2 or January 2022.",
  "decision_basis": "Span confirms the launch date but says nothing about L2 arrival, so the second half of the claim is unsupported."
}

Failure Modes

• Inferential leniency — model marks a claim "fully_supported" when the span only loosely implies it ("rocket launched" → "spacecraft reached destination"). Detect by spot-checking a sample where supporting_phrase does not contain the claim's key noun phrases. Mitigation: emphasize "no additional inference required" and add a few-shot showing a partial-support case.
• Surface-string matching as proof — model treats verbatim word overlap as support, even when the span uses the same words in a different sense. Detect by checking that supporting_phrase is a meaningful clause, not a single word echo.
• Claim atomicity assumption — when the input claim is actually two claims joined by "and" (as in the example), the model may verdict the whole as "not_supported" instead of "partially_supported". This is upstream's job to fix: split conjunctions into atomic claims before calling this card.
• Citation flipping — model occasionally emits "supported" when the span CONTRADICTS the claim. Detect by sampling not_supported and fully_supported outputs and looking for clear contradictions mislabeled as support.

Tuning Notes

• 模型差异：strong 模型（GPT-4 / Claude Sonnet+）在 partial vs full 区分 上稳定；中档模型容易把 partial 误判为 full。可考虑跑两遍取 max strictness（即 partial 优先于 full）。
• 温度：0.0，verdict 稳定性优先。
• 输入要求：claim 应该已经被 atomic 化（一个事实=一个 claim）。 conjunction（"X and Y"）类的复合 claim 在送本卡前先拆。
• 与 rag/answer-grounding-checker 的关系：本卡是 per-claim 的微观审计 （需要明确的 citation span）；answer-grounding-checker 是 per-answer 的宏观审计（不需要 explicit citation，直接对照 retrieved context）。
• 用作训练信号：fully_supported 比例可以作为 RAG 系统 citation precision 指标；建议把 not_supported 样本送人工抽检以校准。

Changelog

• 0.1.0 — Initial card.