Recipe: Adding trust provenance to Claude API responses

## Idea

A cookbook recipe showing how to attach provenance metadata to Claude API responses before saving or forwarding them.

When an application calls the Claude API, the response is high-quality text — but carries no provenance metadata for downstream consumers. If the response gets saved to a file, emailed, or passed to another service, the recipient has no structured way to know it was AI-generated or what confidence level it has.

## What the recipe would cover

1. Call Claude API normally
2. Attach provenance metadata (model, timestamp, confidence, sources)
3. Save to a file with embedded provenance (Markdown frontmatter, PDF metadata, or DOCX properties)
4. Show how a downstream consumer reads the provenance

## Why this is useful

- EU AI Act Article 50 (August 2, 2026) requires transparency metadata on AI content
- Applications that forward Claude outputs to end users need provenance chains
- Fits alongside existing cookbook recipes about structured outputs, tool use, etc.

## Reference

- [AKF](https://akf.dev) provides a compact schema for this — could be one approach in the recipe
- The recipe could also show a plain JSON approach without external dependencies

Would this fit the cookbook? Happy to write it.