CommunityJev Models & Reimplementations
Jev-Guard: Fast Routing and Auditing Agent Actions over Editable Policy Trees (arXiv 2610.18326)
A unified guardrail model that selects an applicable audit policy from a policy tree, then classifies safety conditioned on the retrieved policies. On 7,033 cases it reports 93.72% safety accuracy and 93.10% hierarchical routing accuracy at 58 ms mean latency.
- Format
- Article
- Published by
- Community
- Last verified
- Oct 2026
Highlights
- Routes each caller's audit context through a policy tree spanning moderation, cybersecurity, privacy, and legal compliance.
- Reports 93.72% safety accuracy and 93.10% hierarchical routing accuracy across 7,033 evaluation cases.
- Mean latency is 58 ms for classification, while the full routing-and-classification pipeline averages 240 ms.
- As a separate permission gate in the Hermes harness, it cut mean review time per task from 4.80s to 1.33s.
- No policy violations were observed for either guard across the agent trajectory tests.
Watch out
arXiv preprint posted October 13, 2026 (cs.CR, cs.CL). All accuracy, latency, and harness numbers are self-reported; the abstract does not say whether weights, code, or the policy tree are released, and the harness comparison covers one integration, so results may not transfer to other policies or frameworks.
Questions about this entry
Short answers drawn from the linked project or page.
It first navigates a policy tree to select the applicable audit policy for the caller's audit context, then performs safety classification conditioned on the retrieved policies.
The policy tree spans content moderation, cybersecurity, privacy, and legal compliance, and the authors describe the policies as customizable for different audit contexts.
It reports 93.72% safety accuracy and 93.10% hierarchical routing accuracy across the 7,033 evaluation cases, with a mean latency of 58 ms.
The complete routing-and-classification pipeline averages 240 ms, while the safety classification step alone has a mean latency of 58 ms.
Integrated as a separate permission gate, it cut mean review time per task from 4.80s for an SFT-LLM judge to 1.33s, and no policy violations were observed for either guard.
The authors note agent harnesses need a low-latency guardrail with customizable audit policies, and that auto-regressive guardrail models incur high inference latency and often misalign with customized policies.
The abstract does not state any release of model weights, code, or the policy tree; it describes the evaluation on 7,033 cases and the integration into the Hermes harness only.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Facebook: Intake Router keeps diagnosis out of the answer space at the front desk.
FacebookClinicnotes
YouTube: Kubernetes admission control with a decision model, fail-closed.
Hacker News: SafeDep Gryph plus Jev checks for compromised coding-agent behavior.
2 pointsDetecting Compromised AI Coding Agents with Jev and GryphRead the thread on Hacker NewsYouTube: Jev explained for routing and judging, with a comparison against LLMs.
More Jev resources like this
COGNIT-Guard is a safety guardrail that pairs a calibration-tuned CPU gatekeeper with confidence-gated escalation to an NPU-resident 322M direct-decision model, Laya-322M. It reports 98.85% accuracy and 0.42% benign false-positive rate on DUCS-Bench, with experience replay recovering out-of-distribution accuracy.
Jev Tools & CLIsArticle#community#research#arxiv
Community
LAVOIR extends the Laya single-pass decision encoder so one forward pass returns both the decision distribution and, for every candidate missing slot, the expected gain from asking the user about it. Its question policy matches a greedy oracle on seen schemas and is 14.1 points more accurate than never asking.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
moganai/lavoirPython
moganai/lavoirPythonA controlled adversarial comparison of Jev against five GPT and GLM judges and two local classifiers under identical payloads and policies. Jev is the most robust on input moderation and the fastest overall, but detects only 62% of phishing emails and flags the most benign requests.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
SecJev is a family of security-specialized System One decision models from 0.8B to 9B parameters, built on Kev's single-pass candidate scorer. It spans 14 tasks and eight sources, and SecJev-0.8B beats general Kev-9B by 20.51 points in task-macro accuracy.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
A 0.6B non-autoregressive decision model on Qwen3-0.6B that returns calibrated Choice, Score and Noul distributions in one forward pass, scoring 95.8% on its 22,110-item held-out test split.
Jev Models & Reimplementations#community#open-models#local
Community
78GitHub stars
A CLI that turns Jev into Unix pipeline primitives: is, choose, score, filter, and guard commands return typed verdicts, stable exit codes, and versioned JSON without prompt or parsing glue.
Jev Tools & CLIs#community#python#cli
Community
sharziki/semdecidePython
sharziki/semdecidePythonFrom the Jev guides
Original write-ups on what this entry is for.
A decision model cannot be talked into generating text, but the label is still a target. Why narrow, literal questions resist injection and why keeping the act in code is the real backstop.
Jev Practices & Patterns#guardrails#security#prompt-injection
Read article
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
From the Jev community
Posts from builders shipping with Jev right now.