Hi everyone,
I am an independent researcher seeking an arXiv cs.AI
endorsement for my first preprint.
Paper: “Policy-Based Agentic Systems (PBAS): Where the Model
Proposes and the System Governs”
The paper introduces PBAS — an architectural framework that
formally separates planning from execution in LLM-based agent
systems. The central contribution is a deterministic Policy
Engine that evaluates every proposed plan against an
organizational policy matrix before any execution begins.
Permissions are enforced at the API layer, not via prompt
instructions — making the system structurally resistant to
prompt injection.
The reference implementation (DAF) is fully open-source:
v0.1.0 — Apache 2.0 — 561 tests passing including 53
adversarial security tests.
Happy to share the PDF on request.
Endorsement code: 8TAABS
Thank you for considering — any help is greatly appreciated.
Tony Ochinang
tony.ochinang@tonyos-lab.org