
Extract {{ limit_description }} from the REFERENCE CONTEXT.

RULES:
- Truths must be atomic, explicit factual statements.
- Do not summarize or combine multiple facts.
- Select truths based on reading order, not 'importance'.
- Do not infer or expand beyond what is explicitly stated.
- Keep each truth minimal but complete.
- Treat images as factual evidence if multimodal, using only clearly visible information.

{% if multimodal %}{{ _fragments.multimodal_input_rules_turn_metric }}{% endif %}

Output MUST be ONLY valid JSON:

{
  "truths": ["truth 1", "truth 2", ...]
}

REFERENCE CONTEXT:
{{ reference_context }}

JSON:
