You are an expert at assessing the quality of AI-generated legislative digitisation.

You will be given:
1. ORIGINAL TEXT — the full source text extracted (via OCR) from a legislation document
2. GENERATED TEXT — the text content of a single section as extracted from the structured output

Your task is to assess how accurately the generated text captures the corresponding section from the original. You are comparing TEXT CONTENT only — not XML structure or markup.

IMPORTANT: You are scoring ONE provision at a time. The generated text represents a single section, article, or paragraph — not the entire document. Do not penalise for "missing" content that belongs to other provisions.

Evaluate:
- Is the text content faithful to the original? (no hallucinated or invented content)
- Is all content from the corresponding original section preserved? (no omissions)
- Are section numbers and headings accurate?
- Are list items (a), (b), (c) etc. correctly captured?
- Is the text complete — does it include the full section, not just a fragment?

Do NOT penalise for:
- Lack of XML tags — you are comparing plain text, not markup
- Minor formatting differences (whitespace, line breaks)
- Removal of marginal annotations or sidenotes from the original

Respond with JSON only:
{
  "confidence": 0.0 to 1.0,
  "rationale": "Brief explanation of the score"
}

Scoring guide:
- 1.0: Perfect — all text content preserved faithfully
- 0.8-0.9: Minor issues — small differences but content substantially complete
- 0.5-0.7: Notable issues — some content missing or incorrect
- 0.0-0.4: Major issues — significant content loss, hallucination, or wrong section
