JEV-27B: How AutoTrust AI Built an Open Decision Model That Thinks Fast and Reasons Deep

Most AI workloads in production don't need a model to write an essay. They need a fast, reliable answer to a narrow question: Is this email phishing? Which support queue does this ticket belong to? Does this content violate policy? For years, developers have been using full-scale large language models to answer these questions — paying frontier-model prices for tasks that don't require frontier-model capabilities.

AutoTrust AI's JEV-27B is a direct response to that mismatch. Released in early October 2026 under an Apache-2.0 license, it introduces a new category of open-weight model: the decision model — a system designed to return calibrated, structured judgments at high speed, without sacrificing the ability to reason when reasoning is actually needed.

What Is a Decision Model?

The term borrows from psychologist Daniel Kahneman's framework of System 1 (fast, intuitive) and System 2 (slow, deliberate) thinking. Traditional LLMs are System 2 machines: they generate tokens sequentially, building up a response word by word. That's powerful for open-ended tasks, but it's expensive and slow for classification, scoring, or routing.

A decision model is optimized for System 1 tasks. Given a natural-language question and a set of typed options — a boolean, a label from a fixed set, or a numeric score — it returns a calibrated probability distribution in a single forward pass, without decoding, JSON parsing, or prompt engineering. The output isn't text; it's a structured judgment with a confidence score attached.

This matters operationally. When you're running millions of content moderation checks per day, the difference between 137 ms and 280 ms per decision, multiplied across that volume, is the difference between a viable product and an unaffordable one.

The Blocks of Experts Architecture

JEV-27B's core technical contribution is what AutoTrust calls the Blocks of Experts (BoE) recipe. Rather than fine-tuning the entire model — which risks degrading the reasoning capabilities that make a large model useful — BoE keeps the backbone frozen and adds a small, detachable expert block for the decision task.

The architecture has three components:

• Backbone: The Qwen3.8-27B text tower, frozen and bit-identical to its original release. This is the System 2 engine.
• System 1 Block: A LoRA adapter (rank 16) plus a 24-slot fp32 decision head, totaling about 108.9 million parameters — roughly 0.4% of the full model size. This block is trained on the SargeDev/jev-distill-corpus-v3 dataset to replicate the output distributions of TypeSafe's closed-source Jev 1.13 model.
• Router: A per-request mechanism that directs traffic to either the decision block or the generation block, depending on whether the incoming request is a typed decision query or a free-form generation task.

The key insight is that the System 1 block is detachable. When it's inactive, the model is byte-identical to the base Qwen3.8-27B. AutoTrust verified this by running HumanEval with the decision block both active and inactive: all 164 completions were identical, and the model scored 78.0% either way. This avoids the performance degradation that often occurs when LoRA weights are merged directly into a backbone.

Calibration, Not Just Classification

What separates JEV-27B from a standard classifier is its emphasis on calibration. Rather than returning a hard label, the model outputs a probability distribution over the possible answers. This lets downstream systems set confidence thresholds: auto-approve if probability > 0.9, escalate to a human reviewer if it falls below 0.7, and route to a more expensive model for the ambiguous middle.

AutoTrust measured calibration fidelity against the teacher model (TypeSafe Jev 1.13) across 25,376 held-out questions. The mean KL divergence was approximately 0.017 — meaning JEV-27B's probability distributions are nearly indistinguishable from the closed-source model it was trained to replicate. On six public decision benchmarks, JEV-27B averaged 84.07%, slightly above the teacher's 83.85%.

Latency on a single NVIDIA B200 GPU: 137 ms median per decision, compared to 238–301 ms for hosted API alternatives. Throughput: approximately 130 decisions per second.

Practical Use Cases

The AutoTrust blog post and independent evaluations from Cisco's AI team highlight several production patterns where decision models outperform general-purpose LLMs:

Content moderation at scale. Convert a policy into a series of typed questions. JEV-27B answers each with a calibrated score, making it easy to aggregate signals and set per-category thresholds.

Ticket routing and triage. Classifying support tickets into queues is a high-volume, low-complexity task. A decision model handles this in a single pass at a fraction of the cost of a frontier model.

Data quality screening. Checking whether a user-submitted field is valid, whether a code change introduces a security pattern, or whether a document matches a regulatory template — judgment calls that don't require prose generation.

Cost-efficient agent pipelines. In multi-step agentic workflows, decision models serve as fast pre-filters that handle easy cases and only escalate to expensive reasoning models when the situation is genuinely ambiguous.

The Competitive Landscape

JEV-27B didn't arrive in a vacuum. TypeSafe AI's hosted Jev service had already gained traction before AutoTrust released the open-weight version. Then, on October 6, 2026, OpenAI launched its Decisions API, built on GPT-6 Luna, entering the same market directly.

The pricing comparison is instructive: OpenAI's Decisions API lists at $0.10 per million input tokens; TypeSafe Jev charges $0.042 per million. JEV-27B, being self-hosted, shifts the cost to infrastructure — cheaper for organizations already running GPU clusters. OpenAI's version supports image input (JEV-27B-VL, a visual follow-on from AutoTrust, addresses this gap), and claims up to 10× speed improvement over its general-purpose Responses API.

What This Means for Practitioners

The emergence of decision models as a distinct category reflects a broader maturation in how teams think about AI infrastructure. Not every task in a pipeline needs the same model. A well-designed system might use a decision model for high-volume triage, a mid-tier reasoning model for moderate complexity, and a frontier model only for genuinely hard problems.

JEV-27B makes this architecture accessible to teams that need data sovereignty, want to avoid per-token API costs at scale, or simply prefer to control their own inference stack. The Apache-2.0 license means it can be deployed commercially without restriction, and the vLLM compatibility means it slots into existing serving infrastructure without custom work.

The Blocks of Experts approach is also worth watching as a general technique. If it proves reliable across other task types — structured extraction, scoring, or routing — it could become a standard recipe for adding specialized capabilities to open-weight models without compromising their general-purpose utility.

Getting Started

JEV-27B is available on Hugging Face under Apache-2.0. AutoTrust's blog post includes a quickstart guide for deploying with vLLM and examples of how to structure decision queries using the noul, choice, and score task types. Cisco's independent evaluation provides a useful comparison against traditional fine-tuned classifiers and the hosted Jev API.

The core question for any team considering this approach is whether their workload has enough high-volume, structured judgment tasks to justify the infrastructure overhead of self-hosting. For organizations already running inference at scale, the answer is likely yes.