Sachin Gupta’s AgentTrustCI paper, submitted to arXiv on 4 November 2026, describes a Java/JUnit 5 framework for checking large language model (LLM) agents during software builds. It targets application behavior: whether an agent selected an allowed tool, based its answer on tool output and followed a required workflow. The abstract argues that infrastructure security does not establish these properties.[1]

AgentTrustCI expresses those expectations as policy, grounding and trajectory checks over recorded traces. A deterministic replay implementation supports repeatable tests, while a judge interface allows optional semantic evaluation. The abstract describes a grounded answer reached through a disallowed sequence as an example of why workflow checks add coverage.[1]

The author reports that the full pipeline correctly classified all 55 synthetic cases across seven enterprise domains: 30 injected faults and 25 passing controls, including ten non-trivial controls. It missed none of 12 designated critical cases. Each single-layer configuration classified 35 cases correctly; the corpus deliberately isolates ten faults per layer. These are reported research results, and the abstract explicitly limits them to coverage of the authored regression suite rather than production accuracy.[1]

[1]: AgentTrustCI: Build-Time Policy, Grounding, and Trajectory Tests for Java LLM Agents, Sachin Gupta, arXiv, 4 November 2026.

Sources

  1. AgentTrustCI: Build-Time Policy, Grounding, and Trajectory Tests for Java LLM AgentsarXiv · 4 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

2f87283b5dd9a7dcdbf93000e4d7dd13b99580cda82546262b806e71616ec4fc