The author of an open, model-agnostic format for AI decision provenance is seeking community feedback on a new standard called "interaction cards." This initiative aims to address the gap in verifying how specific outputs were produced, complementing existing model cards that document the model itself.
- A "Causal Seal" attaches a record to a single output, binding the output, causal parameters (model identity, constraints, context, timestamp), and emission moment into a SHA-256 fingerprint.
- Verification is possible with no access to the model or private systems.
- The specification includes a JSON Schema, conformance vectors, and a reference Python verifier installable via pip.
- A gateway emitting seals per response and a browser verifier are available.
- The project is licensed under CC BY 4.0 and has a DOI record on Zenodo.
The author asks whether there is interest in this format within the Hugging Face ecosystem and how it might fit alongside model cards.