Abstract
This paper develops a bounded execution-assurance framework for tool-using AI agents by coupling observation-bounded certification, typed pre-effect enforcement, claim-dependent evidence semantics, and a source-audited reference implementation.
It formalizes when execution properties can be certified from an observation boundary, scopes runtime safety guarantees to mediated committed effects, distinguishes cryptographic integrity from semantic truth, and evaluates the resulting assurance boundaries against the pinned AegisRun / AEGIS implementation.
Research objective
The central problem is not whether a model can propose an action, but whether the surrounding execution system can justify that the action was authorized, checked, mediated, committed correctly, and supported by evidence with clearly bounded meaning. The paper separates what can be certified from observable runtime boundaries from claims that require stronger assumptions.
Resources
The full manuscript PDF is mirrored on this research site for scholarly discovery. The canonical archival preprint version is preserved on Zenodo under DOI 10.5281/zenodo.23119086; its Zenodo version-family DOI is 10.5281/zenodo.23119087. The separate reproducibility artifact has version DOI 10.5281/zenodo.23118665 and version-family DOI 10.5281/zenodo.23118664. The reference implementation is available as AegisRun / AEGIS.
Citation
Saridakis, Stamatis-Christos. “Agentic Execution Assurance: Observation-Bounded Certification, Runtime Enforcement, and Tamper-Evident Evidence.” 2026. DOI: 10.5281/zenodo.23119086.