I’m the author of an open, work-in-progress format for the provenance of AI
decisions, and I’d love feedback from this community — model cards are, in a
sense, the ancestor of what I’m building, and HF champions open, model-agnostic
documentation.
The gap. Model cards document the model. There’s still no open way to
verify how a specific output was produced.
The idea. A “Causal Seal” attaches to a single output a small record that
binds — into a SHA-256 fingerprint — the output, the causal parameters that
governed its generation (model identity, constraints, context, timestamp), and
the moment of emission. Anyone can verify it with no access to the model or
any private system. It’s open (CC BY 4.0) and privileges no vendor. Think of it
as the per-decision, verifiable counterpart to a model card — an “interaction
card.” Complementary, not competing.
What already exists (public, reproducible):
pip install causal-seal
) and a browser verifierMy questions:
Any feedback is welcome — this is early, single-maintainer, open work.