JEV-27B-VL

# Fix list: JEV-27B-VL

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/jev-27b-vl, the October 2026 research run, assessed 21 October 2026. Grade C, 59.5 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on JEV-27B-VL: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Reach, 50 out of 100, up to 19.2 more on the total

Why it scored 50: Memory 10 of 30, the only build is BF16 at 56.0 GB with the adapter and vision encoder, 67.2 GB with a fifth added, which fits the reference machine but nothing smaller. Runtimes 15 of 25, released vLLM, SGLang and Transformers load the architecture with its LoRA, while llama.cpp, Ollama and LM Studio have no route for the adapter. Builds 0, AutoTrust publishes no quantised build of this model. Terms 15, Apache 2.0 with nothing added. Download 10, not gated. Hosted 0, no provider or API serves it.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach):

Whether an agent can run it, and where.

- 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above.
- 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most.
- 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist.
- 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms.
- 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request.
- 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it.

When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach.

## 2. Risk, 45 out of 100, up to 16.9 more on the total

Why it scored 45: File safety 15 of 25, safetensors only, but the documented quick start runs the repository's serve_decide.py. Provenance 15, the files are in the organisation AutoTrust's own demo links to, with a checksum on every file. Safety evaluation 0, none published. Agent guidance 5, general cautions about thresholds and computer use, nothing on prompt injection. Licence risk 10, Apache 2.0 throughout the chain, less 5 because the labels were distilled from TypeSafe's Jev against terms that forbid it. Security programme 0, no way to report a flaw was found.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk):

What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk.

- 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template).
- 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file.
- 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none.
- 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none.
- 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0.
- 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public.

When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk.

## 3. Agent interface, 80 out of 100, up to 3.1 more on the total

Why it scored 80: Read as a decision model, by what it takes and returns. Chat template 10 of 20, the repository's template serves only System 2 and the decision prompt lives in code. Tool-calling format 25, the request and response of POST /v1/decide are documented with the server to run. Trained tool calling 10, one question per request. Structured output 15, typed probabilities are the whole interface. Reasoning control 10, thinking off, auto or on per request with a budget and the reasoning text returned on request. Context 10, 262,144 tokens native.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface):

What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how.

- 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run.
- 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it.
- 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time.
- 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how.
- 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back.
- 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down.

## 4. Documentation and upkeep, 92 out of 100, up to 1.2 more on the total

Why it scored 92: Model card 17 of 20, everything but a knowledge cut-off. Training data 20, the adapter's corpus is published with its sources and sizes, though the base's data isn't. Last change 20, the card changed 7 days ago. Responsiveness 20, AutoTrust answered 5 of 7 questions. Version policy 5, a first release. File health 10, the files agree with each other and the card and no broken-file report waits.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep):

- 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing.
- 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none.
- 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older.
- 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none.
- 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note.
- 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Whether released vLLM (without serve_decide.py) gives the same System 1 probabilities as the development build the card tested, and whether --mamba-cache-mode exists in a release, couldn't be checked without running it
- llama.cpp's LoRA conversion wasn't checked for an adapter that targets lm_head and the linear-attention projections, so llama.cpp, Ollama and LM Studio aren't credited
- 2,019,618 downloads in three weeks with the 30-day and all-time figures equal is far above the text sibling's 269,687, and the founder may want to treat the count as unverified, as BATCH.md does for JiRack
- Licence risk takes 5 off for training on labels distilled from TypeSafe Jev 1.13, whose Master Customer Agreement forbids distillation, read on the evidence recorded. Whether those terms bind OpenRouter users or anyone downstream is a legal question for the founder, and a negative event would need TypeSafe to object

## Weaknesses

- Trained on labels distilled from TypeSafe's hosted Jev, whose customer agreement forbids distillation
- No quantised build from AutoTrust, so about 67 GB is the floor
- The documented server runs the repository's serve_decide.py and needs a vLLM development build
- One question per request, and image decisions are zero-shot with calibration unmeasured
- No safety results, prompt-injection guidance or security contact

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Quickest start is `hf download autotrust/JEV-27B-VL` then `bash JEV-27B-VL/serve.sh`, which serves POST /v1/decide on port 8000, but it runs the repository's serve_decide.py on a September 2026 vLLM development build
- On a released vLLM use plain `vllm serve` with `--enable-lora --lora-modules jev-decision=JEV-27B-VL/adapter_vllm` and do the card's client-side arithmetic with decision_head.json and calibration.json
- Keep `--max-num-seqs 8`, the card says larger batches return wrong System 1 probabilities on vLLM's multimodal LoRA path
- When calling /v1/completions or /v1/chat/completions yourself pass top_k 0 and top_p 1.0, or generation_config.json's top_k 20 and top_p 0.95 truncate the probabilities
- Choices beyond 16 options use labels the head never saw in training, and the response's adaptation field says wide-labels
- Downscale images to about 448 px to keep vision tokens and latency down
- Treat it as AutoTrust's model, not TypeSafe's, and its request shape differs from TypeSafe's /v1/systemone

## What the review panel asked for

- target a released vLLM
- state a version policy
- a 4-bit build
- several questions per request
- a security contact
- state the distillation terms

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.