Jev-Guard: Fast Routing and Auditing Agent Actions over Editable Policy Trees

An agent reads external documents or tool responses that may contain malicious instructions.

What you would change

Replay attacks and legitimate lookalikes in an isolated copy of your agent workflow.

Replay attacks and legitimate lookalikes in an isolated copy of your agent workflow.

Compare against

Your existing checks with the same tasks, permissions, and tool responses.

1. Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.
2. Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.
3. Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.
4. Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test.

What to measure

Successful attacks missed, legitimate actions blocked, detection delay, and review cost.

When to keep it

Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation.

recommendation

test

research area

safety and guardrails

technical change

Policy-tree routing with unified guardrail classifier

published

2026-10-13

A small non-autoregressive guardrail fits a single 24 GB GPU and directly reduces agent-audit latency, but the claimed accuracy and zero-violation results come from one harness and one evaluation set, so adoption should follow validation on the reader's own policies and agent workloads.

1. claim 01Paperraft inference
   Read the brief interpretation

   Replaces an autoregressive guardrail model or LLM-as-judge used for agent action auditing with a single small model that first routes over an editable policy tree and then classifies safety conditioned on the retrieved policy. Costs include training or obtaining the specialized model, maintaining the policy tree spanning moderation, cybersecurity, privacy, and legal domains, and a residual error rate of roughly 6% on both routing and classification; integration is demonstrated only within the Hermes harness. Can fail through misrouting to the wrong policy branch (which invalidates the downstream audit), residual misclassification of unsafe actions, and policy drift if the editable tree is not kept synchronized with actual deployment rules.

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source
2. claim 02Paperraft inference

   Permission-gate review time falls from 4.80s per task with an SFT-LLM judge to 1.33s (about 3.6x); the model itself reports 58ms mean classification latency, 240ms for the full routing-plus-classification pipeline, 93.72% safety accuracy, and 93.10% hierarchical routing accuracy on 7,033 cases.

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source