Causal Seal — an open format for decision provenance in generative AI

Status: v1.0-draft — published for community review. Site: causalseal.org · Paper: doi.org/10.5281/zenodo.21431267

Content provenance (C2PA) proves where an artifact comes from. Log integrity proves a record wasn't altered. The Causal Seal proves something neither does: why a generative system produced a specific output — by cryptographically binding the output to the causal parameters that governed its generation.

• SPEC.md — the specification (data model, canonicalization, verification, conformance, regulatory mapping).
• causal_seal.py — minimal reference implementation, standard library only: emitter + Level-1 verifier + CLI (demo, verify, selftest).
• causal-seal.schema.json — JSON Schema for automated validation.
• test-vectors/ — computed vectors: a valid seal (with its exact expected fingerprint) and a tampered one that must fail.
• docs/verify.html — public verifier (live): paste a seal (and optionally the output text), get 🟢/🔴 — computed entirely in the browser, nothing sent anywhere. Validated against the test vectors.
• paper/PAPER.md — companion paper (draft): Causal Seals: Decision Provenance for Governed Generation · archived preprint: 10.5281/zenodo.21431267.
• pyproject.toml — packaging: ready for pip install causal-seal at publication.

$ python causal_seal.py selftest
OK   invalid-001.json: verifier says FAIL (fingerprint mismatch) — expected FAIL
OK   valid-001.json:   verifier says PASS (seal verified, Level 1) — expected PASS
selftest: ALL GOOD

The specification standardizes the envelope of proof, not the engine of governance: implementing the format is intended to be royalty-free (see SPEC §9); how an emitter produces a genuinely causal state is its own concern — and its own advantage.

Spec text: CC BY 4.0. Reference implementation: zero dependencies, Python 3.10+.