Problem Statement
When one LLM passes output to another (e.g., summarization, transformation), metadata is lost. There’s no mechanism to preserve or verify origin across multiple LLM generations.
Proposed Solution
- Enable metadata passthrough and relaying between models.
- If Model A creates a text, and Model B modifies it, metadata includes both origins with timestamps and hashes.
- Maintain provenance chain with nested or chained signatures.
Alternative Solutions
- Flatten metadata to “last model only” (loses history).
- Manual logging of provenance (non-standard, easy to omit).
Use Cases
- A content platform chains summarization, translation, and rewriting steps through multiple LLMs and wants to preserve full provenance.
- An enterprise uses one LLM to draft, and another to edit – both steps need attribution.
Implementation Ideas
- Chain metadata with a signature tree.
- Use canonical ordering to preserve integrity and prevent tampering.
- Update encoder to optionally accept prior metadata and extend it.
Additional Context
Could become essential in regulated or high-trust AI pipelines.