New Framework Ensures Provenance Integrity in AI Agent Workflows

In institutional settings, simply achieving a "correct" outcome from an AI agent is insufficient; the process must also be auditable and verifiable. This research introduces the concept of "governed execution" for agentic workflows, which ensures that all decisions, completions, and responses to change are backed by inspectable provenance. The paper presents Matrix, a deterministic causal-state layer designed to implement this. Matrix records authority and fact dependencies, rigorously verifies completion evidence, and intelligently invalidates only the affected parts of work when changes occur. Comparative studies showed that while governed and direct workflows often reached the same final outcomes, only the governed path consistently maintained governing evidence, prevented unsupported closures, and limited recovery efforts to truly dependent tasks. A specific challenge involving role-separated transfer highlighted that a deterministically enforced completeness contract could over-block work produced outside its authoring context, indicating areas for refinement. Matrix's primary role is to provide an institutional integrity layer, making agentic work auditable and independently verifiable.