Causal Seals: Decision Provenance for Governed Generation

Two families of standards address trust in AI-produced artifacts: content provenance (C2PA) proves where an artifact comes from, and integrity protocols prove that a record was not altered. Neither answers the question that regulators have now made statutory for high-risk AI systems: why did the system produce this specific output? We introduce decision provenance: a verifiable binding between a generated output and the causal parameters that governed its generation. We describe (i) governed generation, an architectural pattern that decouples deterministic control from stochastic text execution; (ii) the Causal Seal, an open, model-agnostic format that binds output, causal state and emission time under a recomputable SHA-256 fingerprint; and (iii) a zero-dependency reference implementation with computed test vectors, alongside a production deployment that has sealed every generated response since mid-2026. We map the format onto the record-keeping obligations of the EU AI Act (Articles 12 and 19), enforceable for high-risk systems as of 2 August 2026, and discuss what a seal does and does not prove. The specification, schema, vectors and implementation are released for community review.