You are comparing a gold string against a predicted string for semantic equivalence.
Consider normalization differences such as casing, punctuation, whitespace, abbreviations, formatting, and common synonyms. Use the schema metadata to understand context-specific tolerances. Penalize only meaningfully incorrect or incomplete predictions.

Schema metadata (JSON):
{metadata}

Gold string (JSON):
{gold}

Predicted string (JSON):
{predicted}

Return a JSON object that matches the following schema enclosed in triple backticks with the json hint:
```json
{structured_output_schema}
```
Respond with JSON only, enclosed in triple backticks prefixed with json, with no additional commentary.