Artifact for 'Deon: Runtime Guards for Policy-Safe Agentic Service Composition under Attack'

Deon guards actions proposed by an agent before services execute them. A policy automaton checks permissions and obligations, and a calibrated detector addresses content violations. This release contains the implementation, generated benchmark code, experiment drivers, recorded results, and figure and table scripts for the ICSOC 2026 paper. The local checks verify package integrity and regenerate the core and stress experiments; optional Ollama and AgentDojo runs require separate model downloads and dependencies. The code archive is the release published on GitHub (see Related works). The accompanying six-page description gives installation commands, expected outputs, seeds, resource needs, and a claim-to-result map. The stored language-model summaries are single stochastic runs. Original model plans, AgentDojo episode logs, per-call refusal logs, and immutable model digests are unavailable, so those interactions cannot be reconstructed exactly. Rights note: Code and configuration: Apache-2.0. Original documentation, fixtures, and processed results: CC-BY-4.0 unless marked otherwise. The included paper PDFs retain publication terms. Dependencies, Ollama, AgentDojo, and model providers retain their own licenses; no model weights or AgentDojo code are redistributed. See LICENSES/ and THIRD_PARTY_NOTICES.md inside the code archive.