<task>
You are an expert annotator. Identify every substring of a generated **answer** that is a hallucination with respect to the given **source**.

## Grounding rule (read first)
- Judge the answer **only** against the <source>. Treat the source as the single source of truth.
- A claim not stated or entailed by the source is a hallucination **even if it is true in the real world**. Never use outside or prior knowledge to support a claim.

## Language
- The source and answer are written in **${lang}**. Respond **in ${lang} only**.

## What IS a hallucination
A substring that:
- **contradicts** the source (a different number, date, name, entity, or relationship), or
- **adds** a fact, cause, detail, or quality the source does not state or entail, or
- **references** a named element (person, place, identifier, section, attribute) absent from the source, or
- **overstates** the source (e.g. source says "may help", answer says "cures").

## What is NOT a hallucination
- A statement consistent with the source, including a faithful paraphrase, rounding, reordering, or recombination of information that IS in the source.
- Framing or boilerplate that makes no factual claim (e.g. "Here is the answer:"). Do not flag it.
- An explicit refusal such as "Unable to answer based on given passages."

## How to choose spans
- Flag the **minimal** substring carrying the unsupported information; do not pad with surrounding supported text.
- Copy each span **verbatim** from the answer — character for character, including punctuation, casing, and spacing. Do not paraphrase, translate, fix typos, or normalize. A span that is not an exact substring of the answer is discarded.
- If one sentence mixes supported and unsupported content, flag only the unsupported part.

## Steps
1. **Read** the answer inside <answer>…</answer>.
2. **Check each claim** against <source>…</source> using the rules above.
3. **Return** only confirmed hallucinations, each an exact substring of the answer.
4. ${response_format_block}
</task>

${fewshot_block}

<source>
${context}
</source>

<answer>
${answer}
</answer>
