TraceGuard-X: Adaptive Collusion Resistant Monitoring for Agentic AI Systems, a hierarchical governance framework with constitutional dimension prompting…

This paper tackles an important problem in AI control by extending structured monitoring toward adaptive adversaries, lower monitoring costs, and cross-environment transfer. The proposed ideas—including rubric-aware adversary modeling, Constitutional Dimension Prompting, collusion detection, and Pareto-based routing—are well motivated, and the cost-performance tradeoff achieved by the routing policy is a useful practical result. The engineering infrastructure and implementation are also well organized.

The primary weakness is the gap between the paper's claims and the experimental evidence. Several headline contributions, including robustness against adaptive adversaries and zero-shot transfer, are described but not fully evaluated, with results remaining preliminary. The paper would be significantly stronger by focusing on one core mechanism, validating it against the proposed adaptive attack model, and reporting quantitative transfer results. Aligning the abstract and conclusions with the current experimental evidence would also improve the paper's credibility.

Overall, this is a promising systems paper with strong conceptual framing and implementation, but its central claims require more comprehensive experimental validation.

Extending multi-dimensional monitoring to combat collusion is a vital direction for agentic AI control. The breadth of mechanisms introduced, such as Temporal Graph-of-Thought Monitoring and Dynamic Relevance Routing, is impressive. Demonstrating that a three-dimension routing policy achieves 84% of full-ensemble AUC at a 60% API cost provides great practical value for deployment. However, the presentation feels slightly dense due to the high volume of acronyms and subsystems. Completing the pending full adaptive adversarial evaluation will be critical to proving the system's empirical robustness.

To provide an overview of a very difficult challenge in AI control, this research provides a method of monitoring that can be reliable even if the monitoring agent (the controller) and the monitored agent (the controlled entity) are adversarially aligned.

The number of different approaches presented here are numerous, especially given this was completed by one person over a single weekend. There were many innovative concepts presented here. Some examples include, using Constitutional Principles to limit the actions of the Monitor, instead of limiting the actions of the Agent. Using a Runtime Kolmogorov-Smirnov Divergence as a Collusion Signal. And Pareto Cost-Routing of Monitoring Dimensions. The author provided a model of honesty about their current understanding of their results in the Limitations Section, as well as providing an honest assessment of what they do know through the Hypotheses Appendix.

Some possible suggestions for improving this work:

(1). Make sure the Abstract and Conclusion accurately reflect the results of your actual experimental results. The major statement made in the Abstract regarding Robustness is based upon a Smoke Test (Direction 1) with only 2 iterations of testing and does not represent the complete experimental results. Although the authors note that the complete evaluation will occur after additional funding has been obtained, it represents the strongest statements of fact based upon experiments that have not occurred. In order to improve this work either perform a reduced evaluation of the Adaptive-Adversary Evaluation or clearly identify these mechanisms as proposed and untested.

(2). Increase the size of your Calibration Set. The values reported (Cohen's d > .96, p < 1E-05, and AUC = .98) were generated using a total of 40 trajectories per Dimension (20 Benign and 20 Attack Trajectories). Due to having only 20 trajectories per Class, I believe your Statistics to be Fragile and Optimistic. Having at least 100 trajectories per Class (which you mentioned in Future Work), would increase the credibility of your Separation Claims, as well as require you to use a Held-Out Test Set that is independent of your Calibration Set in order to report AUC.

(3). Provide Quantitative Evidence of Transferability. Your discussion of zero-shot transferability to Infrastructure-as-Code, Code Generation, and Web Agents states that there were promising preliminary results. Providing at least one quantitative example of transferability for each Environment would support your HiDAgT Story.

(4). Protect Against Failure Modes of the Monitor. As several Dimensions utilize Reasoning Capabilities of Large Language Models, I suggest reporting how frequently the Judge hallucinates or reverses its decision under Prompt Changes. Additionally, I recommend evaluating how Federated Median Aggregation performs when a Majority of Judges are Compromised versus a Minority.

Overall: a solid foundation of ideas and architectures; however, in order to further develop this body of knowledge into actionable research, we need evidence as most of your major contributions are currently architectural/hypothetical in nature, whereas most of your claims have not been validated experimentally.