Matrix: Ensuring Provenance Integrity in Agentic Workflows
A recent paper on arXiv (2608.12761) presents Matrix, a deterministic causal-state layer aimed at ensuring provenance integrity within agentic workflows. The authors contend that merely assessing workflows based on their outcomes is inadequate in institutional environments, where actions might depend on incorrect authorities, unverified completion claims, or outdated information. Matrix captures authority and fact dependencies, authenticates completion evidence, and selectively nullifies affected tasks. In controlled experiments, both governed and direct workflows achieved similar results; however, the governed approach consistently maintained governing evidence, rejected unsupported conclusions, and restricted recovery to dependent tasks. A role-separated transfer challenge demonstrated that a deterministically enforced completeness contract excessively blocked synthetic packets produced outside the governed context. The paper falls under the 'new' announcement category and is accessible on arXiv.
Key facts
- Paper ID: arXiv:2608.12761
- Title: 'Correct Is Not Governed: Provenance Integrity in Agentic Workflows'
- Introduces Matrix, a deterministic causal-state layer
- Matrix records authority and fact dependencies
- Verifies completion evidence and selectively invalidates affected work
- Governed workflows preserved governing evidence, refused unsupported closure, and limited recovery
- Role-separated transfer challenge over-blocked synthetic packets
- Published on arXiv with announcement type 'new'
Entities
Institutions
- arXiv